Privacy computing visual sensing unit and output desensitization visualization method thereof

Abstract

The invention discloses a privacy computing visual sensing unit and a desensitization visualization method for output of the privacy computing visual sensing unit. The sensing unit is formed by integrating an optical signal acquisition module, an edge calculation module and a data output module, and an event-type image sensor is adopted to asynchronously acquire pixel-level illumination change and output an original event stream. The edge calculation module locally converts an event stream into a privacy calculation visual data stream containing body part semantics, readable expression of the motion state is achieved by distributing distinguishable visual features to different parts, and image reconstruction or artificial intelligence recognition is not depended on. The whole system does not generate, store and transmit complete image frames in any form, and privacy protection is achieved from the source. The visual data stream can be stored in a computer readable medium or transmitted through a network, and is suitable for scenes with high privacy requirements, such as home-based care for the aged, medical monitoring and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent visual sensing and privacy computing, and particularly relates to a non-imaging behavior observation device based on an event-type sensor and a desensitization visualization method. BACKGROUND

[0002] In the field of intelligent monitoring and behavior perception, traditional video monitoring systems rely on CMOS image sensors to periodically collect image frames, and realize behavior analysis through subsequent algorithms. Although such systems can provide rich visual information, the original images contain sensitive content that can identify individual identities, and there is a serious risk of privacy leakage, making it difficult to promote and apply in high-privacy-sensitive scenarios such as elderly care and home security. In the prior art, some schemes use image blurring, pixelization, or feature extraction to discard the original image for privacy protection, but these methods still need to generate complete images, and cannot avoid privacy exposure at the collection end. Another type of non-visual scheme based on infrared, radar, or pressure sensors has certain privacy, but it is difficult to provide intuitive human motion form information, affecting the accurate judgment of the behavior state by nursing staff.

[0003] In recent years, event-type image sensors have attracted attention due to their asynchronous response, low power consumption, and no image output characteristics. Their output is an event stream of pixel-level illumination changes, which naturally has privacy advantages. However, there are essential differences between the data output mechanism of event-type sensors and traditional image sensors: they respond to illumination changes in an asynchronous manner, and only output events when the pixel state changes, resulting in a highly sparse and bursty data stream. In high-dynamic scenes such as rapid motion, sudden changes in illumination, or multiple target activities, the sensor may produce a large number of events in a very short time, forming a transient data peak. Since the event stream is transmitted through serial interfaces such as MIPI CSI-2, the bandwidth and receiving end buffer capacity are limited. When the burst data volume exceeds the interface throughput capacity or the receiving end processing speed, it is easy to cause communication buffer overflow, resulting in event loss, timing disorder, or transmission delay. Such data overload problems directly affect the integrity of subsequent optical flow reconstruction and motion analysis, and further cause trajectory breakage, false detection, or system response lag, seriously weakening the system reliability.

[0004] Existing applications are mostly focused on using event streams directly as inputs for machine learning models, lacking visualization expression mechanisms for human observation, and failing to fully exploit their potential in human-machine collaborative monitoring scenarios. How to realize readable presentation of human motion state without generating images, while ensuring stable transmission of event streams in high-dynamic scenes, has become a key technical bottleneck restricting the landing of event-type sensing technology in privacy-sensitive scenarios. SUMMARY

[0005] The present application aims to solve the technical problems that the existing visual perception technology is difficult to balance the observable behavior state and the unidentifiable individual identity at the acquisition end. The traditional image sensor outputs complete pictures periodically with a fixed frame rate, which is equivalent to continuously generating image sequences, resulting in individual identity information being recorded at the data source. Even if subsequent processing is performed by blurring, shielding or encryption, the original image may still be illegally extracted or leaked at the acquisition device end, and the essential privacy protection cannot be achieved. Although some non-imaging sensing schemes avoid image generation, they can only provide personnel presence, location information or rough motion judgment, lack intuitive expression of human body movement form, and are difficult to meet the actual application needs of behavior interpretation.

[0006] The event-type image sensor adopts an asynchronous working mode, and outputs event data containing spatial coordinates, change polarity and accurate time stamp only when the pixel-level light intensity changes significantly, without full-frame scanning, which naturally avoids the generation of complete images. The original event stream output by the sensor is a 64-bit binary data packet, containing 10-bit X coordinate, 10-bit Y coordinate, 1-bit brightness change polarity, 6-bit reserved bit and 37-bit microsecond-level time stamp, with the characteristics of high time resolution and low data redundancy. However, the event stream is a sparse space-time data sequence, which cannot be directly observed and understood by human beings. Existing applications often use it as the input of machine learning models, and lack visualization expression mechanism for human eye readability.

[0007] In view of the above problems, the present application provides a privacy computing visual sensing unit and a desensitization visualization method thereof, which realizes closed-loop processing from original perception to privacy output through a highly integrated hardware architecture. The sensing unit is integrated in the same packaging shell at the physical level by a light signal acquisition module, an edge computing module and a data output module, forming a hardware entity that runs independently. The light signal acquisition module includes an optical lens and an event-type image sensor chip, which is used to capture the change of ambient light intensity and output the original event stream. The edge computing module is composed of an embedded processor, a local storage unit and a special processing circuit, which is directly connected with the sensor and performs real-time processing on the original event stream inside the device. In order to ensure the stable transmission of the original event stream in the high dynamic scene, the present application introduces an adaptive communication parameter regulation mechanism based on scene mode in the data processing interface module. The mechanism combines the real-time clock information of the system, and judges the current scene mode based on the activity intensity of the original event stream; according to different scene modes, the communication packet parameters in the data transmission process are dynamically adjusted to realize adaptive packaging and transmission management of the event stream, effectively preventing buffer overflow or transmission delay caused by data burst, and ensuring the continuous and stable operation of the system under the condition of low power consumption and low bandwidth. Through time window aggregation, spatial density clustering and motion consistency analysis, the system identifies the active areas corresponding to different parts of the human body, and maps these areas to semantic visual units based on statistical characteristics. The data output module encapsulates the processing result as desensitization visualization data stream, which is transmitted externally through a standard communication interface.

[0008] The desensitization visualization data stream does not contain original events or image frames, but adopts a lightweight binary format. Each frame of data contains a frame header, a timestamp, a target number, and the number, position, motion vector and part feature array of each target. The part feature array defines the rendering parameters such as the graphic type, color, line width or flashing mode of the head, upper limbs, lower limbs and torso, which supports the restoration of readable graphics such as sparse point cloud, cumulative contour or structured human figure on the terminal device. The whole process is completed locally in the sensing unit, and the original event stream does not go out of the device, so the privacy protection is realized from the source.

[0009] In actual deployment, the present application supports multiple system architectures to adapt to the needs of different application scenarios. In the individualized service environment, a single or multiple sensing units can directly transmit the desensitized privacy computing visualization data stream to a remote server platform through a wireless communication network, including Wi-Fi, cellular network or low-power wide-area network. After secure authentication, the user terminal accesses the remote service platform through the Internet to view the non-visual dynamic behavior representation generated by processing in real time or retrospectively. This direct deployment mode of the remote platform is suitable for scenarios that require remote monitoring and flexible access, and all original event streams are converted by privacy computing locally, without uploading unsensitized data, realizing the principles of privacy protection and data minimization at the perception source.

[0010] In a distributed application scenario, multiple sensing units access a local area network through a wired or wireless network, and the desensitized privacy calculation visual data stream is uniformly transmitted to a local central storage device, including a network attached storage (NAS), an industrial computer or an edge server with data persistence capability. Authorized users can access the storage content on demand through the display terminal in the local area network and review the historical behavior records. The local closed-loop architecture supports centralized management and localized storage of data, sensitive information does not leave the internal network, and meets the control requirements of data sovereignty and privacy compliance.

[0011] In a high-security sensitive application scenario, the system can adopt a hybrid architecture of local and external cooperation. Each sensing unit transmits the generated privacy calculation visual data stream to the local storage device for long-term retention, ensuring controllable data sovereignty. Only when the preset conditions are met, such as accepting external audit, regulatory review or reporting service subsidies, the data upload mechanism is triggered. The uploaded content is aggregated and only contains de-identified statistical indicators, such as activity frequency, bed leaving frequency, service response time, etc. Non-individual granularity summary information is transmitted to the external platform through an encrypted channel. This mechanism supports compliance verification and third-party supervision while ensuring that core behavior data is always isolated within the internal network, achieving a balance between security and auditability.

[0012] The desensitized data stream can be transmitted to a mobile terminal or a fixed display device through Ethernet, Wi-Fi or wireless protocol, and rendered into a dynamic visual interface by the supporting client software, allowing users to remotely observe the human motion state without accessing any image data. The entire system does not generate, store or transmit any form of image frame during the entire process of data acquisition, processing and transmission. The output content cannot be reversed to an identifiable individual identity image, achieving privacy protection from the perception source. The data stream can be stored in a computer-readable storage medium, supporting both local playback and remote access modes.

[0013] The present application realizes the unification of high privacy and high readability through hardware integration and localization. The system can be flexibly configured according to actual needs, supporting remote monitoring, local audit and compliance reporting and other diversified functions under the premise of ensuring data security. This scheme is suitable for all technical fields that require non-intrusive behavior observation and have strict requirements for personal privacy protection, especially suitable for large-scale popularization and application in sensitive scenarios such as smart elderly care, medical monitoring and judicial drug rehabilitation.

[0014] Compared with the prior art, it has significant beneficial effects. By directly converting the raw event stream into desensitized visualization data with part semantics locally in the sensing unit, the image generation link is eliminated from the data acquisition source, preventing the risk of personal information leakage. Traditional monitoring systems generally use the "first image, then software desensitization" mode, even if the image is blurred or obscured, there is still a possibility of restoring the privacy picture through algorithm enhancement, and the system does not generate any image frame from the beginning of sensing, solving this inherent defect.

[0015] The generated visualization data stream is presented in the form of sparse point cloud, dynamic contour or structured graph, only containing the motion trajectory and behavior characteristics of each part of the human body, without involving identifiable information such as face, skin color, clothing, etc. This expression method not only avoids the privacy invasion problem caused by traditional cameras, but also retains the spatiotemporal dynamic characteristics of human motion, enabling nursing staff, guardians or security personnel to intuitively judge behavior states such as falling, wandering and sitting for a long time, making up for the shortcomings of non-imaging devices such as infrared sensors and millimeter wave radars in terms of information expression.

[0016] The system uses edge computing architecture, and all sensitive data processing processes are completed locally, including event clustering, part division and graph generation, without the need to upload raw data to the cloud or central server. This design not only greatly reduces the dependence on network bandwidth and avoids response lag caused by network delay, but also improves the real-time performance and reliability of the system. Even in the case of network interruption, the local device can still run independently and store key events, ensuring the continuity of monitoring services.

[0017] The output data format is specially optimized, using lightweight binary encoding, with a single frame data volume of less than 100 bytes, supporting efficient storage and low-power transmission. This data can be rendered into dynamic human animation on mobile terminals such as mobile phones, tablets or smart watches, making it easy for family members or medical staff to remotely view the activity status of the elderly, promptly detect abnormalities and take emergency measures, especially suitable for home-based care, community care and other long-term remote monitoring scenarios.

[0018] In addition, the entire system does not rely on complex image reconstruction algorithms or large-scale artificial intelligence model training, avoiding misjudgment and performance degradation caused by model bias, overfitting or insufficient computing power. Its core processing flow is based on event-driven deterministic algorithms and lightweight neural networks, ensuring high accuracy while having low power consumption, high stability and low cost advantages.

[0019] The application fuses event type visual sensing and edge intelligent processing technology, and constructs an intelligent sensing system for protecting privacy from the source. The system has technical advantages of surpassing traditional monitoring schemes in data essence, processing mechanism, transmission efficiency and deployment flexibility, etc., not only meets the increasingly strict personal information protection regulations, but also provides safe and reliable technical support for the fields of smart care for the elderly, public security and medical monitoring, etc.

[0020] By introducing an adaptive communication parameter regulation mechanism based on the joint judgment of the system real-time clock and the activity intensity of the event stream, the problem of instantaneous data overload caused by dynamic scene mutation is effectively alleviated. The mechanism enables the data transmission process to dynamically adjust the encapsulation strategy of the communication packet according to the actual load, improving the data throughput efficiency in high activity scenes, while reducing the transmission delay in low activity scenes, guaranteeing the integrity and time sequence continuity of the event stream. Thus, the overflow of the receiving end buffer, event loss or processing blockage is avoided, ensuring the stable transmission of the original event data into the privacy computing engine, providing reliable input for subsequent optical flow reconstruction, motion analysis and behavior recognition. Without increasing the hardware bandwidth or improving the processor performance, the stability and response ability of the system are significantly enhanced, the continuity and accuracy of the visual output are improved, thereby guaranteeing the long-term stable operation of the privacy computing visual sensing unit in complex lighting and multi-target motion environment, and improving the practicability and reliability of the overall system. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is a system architecture and data flow diagram of the privacy computing visual sensing unit in the embodiment of the application, showing the integrated structure and data processing path of the optical signal acquisition module, the edge computing module and the data output module.

[0022] Figure 2 It is a raw event stream data structure diagram output by the event type image sensor in the embodiment of the application, used to explain that each event contains four-tuple data format of X coordinate, Y coordinate, time stamp and brightness change polarity.

[0023] Figure 3 It is a comparison diagram of the raw event stream and the frame type image data in the motion capture effect in the embodiment of the application, used to explain that the event data has the advantages of microsecond level time resolution and no motion blur.

[0024] Figure 4 It is a processing flow diagram of the edge computing module for visual identification of human body parts in the embodiment of the application, used to show the whole process from event stream clustering, spatial analysis to part semantic mapping.

[0025] Figure 5A frame structure diagram of the privacy computing visualized data stream output in the embodiment of the present application, used to illustrate that it contains target number, position, motion vector and part feature array in a lightweight binary format.

[0026] Figure 6 A presentation effect diagram of the privacy computing visualized data in the display terminal in the embodiment of the present application, used to show that the human motion state is expressed in a readable manner by different colors and geometric shapes for part differentiation.

[0027] Figure 7 A comparison diagram of the privacy computing visualized output effect in different application scenarios in the embodiment of the present application, used to show the dynamic recognition capability of the system for standing, walking, taking objects and other behaviors in the room.

[0028] Figure 8 A deployment schematic diagram of the sensor direct connection to the cloud in the embodiment of the present application, showing the whole process that the sensor transmits data to the cloud through WiFi / 4G, and the cloud pushes the de-identified human animation to the user mobile terminal in real time.

[0029] Figure 9 A deployment schematic diagram of the storage in the local area network in the embodiment of the present application, showing the process that the sensor connects the local storage device through PoE, and the user directly accesses the storage device through the local area network WiFi to view the de-identified animation.

[0030] Figure 10 A deployment schematic diagram of the mixed transmission architecture in the embodiment of the present application, showing the dual-channel data flow that the sensor→local storage device→storage device uploads to the cloud through WiFi / traffic card→the cloud pushes to the mobile terminal, realizing the dual protection of local security and remote access. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0032] As shown in Figure 1 The embodiment of the present application provides a privacy computing visual sensing unit, which is composed of an optical signal acquisition module, an edge computing module and a data output module. The three parts work cooperatively to form an independent hardware system, which can express the human motion state in a readable manner without generating a complete image, and realize privacy protection at the data acquisition source.

[0033] The light signal acquisition module is responsible for capturing the change information of light in the environment. This module includes an optical lens and an event-based image sensor. The event-based image sensor works differently from traditional cameras. Instead of taking periodic shots of the entire picture, it responds to changes in the intensity of each pixel in an asynchronous manner. When the brightness of a pixel changes beyond a pre-set threshold (e.g., a log light intensity change of ±15%), the sensor outputs a record containing the spatial coordinates (X, Y) of the pixel, the time of occurrence (timestamp, with an accuracy of microseconds), and the direction of brightness change (i.e., brighter or darker). These records constitute the raw event stream 101. Since data is only output when there is a change, the bandwidth of this data stream is much lower than that of traditional video streams, with a typical rate of between 10,000 and 100,000 events per second.

[0034] The raw event stream 101 is then transmitted to the edge computing module for local processing. The edge computing module uses a low-power embedded processor, such as a chip based on the ARM Cortex-M7 or RISC-V architecture, with a clock speed of no less than 400 MHz, and can integrate an FPGA or a dedicated AI acceleration unit to improve real-time processing capabilities. In this module, the received raw event stream 101 is first subjected to spatial-temporal window filtering, such as setting a 50-millisecond time window to aggregate high-density event points into multiple event active regions. The clustering algorithm can use methods such as DBSCAN or mean shift, and each active region corresponds to a potential moving target.

[0035] After identifying the active regions, the system further analyzes the internal event distribution characteristics, including event density, consistency of motion direction, and spatial relative position relationship, to establish a correspondence with human body parts. For example, the head usually appears as a circular or elliptical region with high event density and smooth motion trajectory; the upper limbs (such as arms and hands) appear as slender linear structures with fast motion speed and variable direction; the lower limbs (such as legs and feet) appear as thicker linear structures with periodic swinging characteristics; the torso is a larger rectangular or elliptical region with smaller motion amplitude. Based on the analysis results, different visual features are assigned to different body parts.

[0036] In terms of visual display, the system supports multiple expression schemes to adapt to different usage scenarios and user needs. For example, it can be chosen to display in grayscale mode, all body parts are represented by white lines or contours, and only through shape differences to distinguish the head, limbs and torso, this mode is suitable for minimalist style or need to reduce visual interference environment. It can also be chosen to dynamically highlight the current moving human target, and the static target is gradually faded or hidden, so as to help users focus on the objects in the activity. In addition, color labeling can also be used to assign different colors to different body parts to more clearly judge the behavior state. For example, the head can be represented by a red dot, the upper limbs by a blue line, the lower limbs by a green thick line, and the torso by a yellow outline frame. This color labeling method helps nursing staff quickly identify limb movements such as lifting hands, walking or falling, and other key behaviors. The selection of all these visual features does not depend on image reconstruction or artificial intelligence recognition models, but is directly generated based on the statistical features of the event stream, ensuring lightweight, interpretable and efficient processing.

[0037] After completing the part recognition and identification, the edge computing module encodes these structured information into a lightweight binary format data packet, forming a visualization data stream 102. The structure of the data packet includes frame header, timestamp, target number, unique identification ID of each target, type, spatial coordinates (X, Y), motion vector (Vx, Vy), and visual feature array of each part. The entire data packet length is controlled between 64 to 256 bytes, suitable for low-bandwidth transmission and efficient storage.

[0038] The generated visualization data stream 102 is then sent to the data output module. This module supports multiple communication interfaces, including wired methods such as Ethernet (RJ45) and USB 2.0, and wireless methods such as Wi-Fi 802.11n, Bluetooth Low Energy (BLE), or 4G / 5G modules. Through these interfaces, the visualization data stream 102 is transmitted to the external display terminal 103.

[0039] The display terminal 103 runs a dedicated application or web client, receives and parses the data stream, and dynamically renders a desensitized human animation. Users can observe the activity state of the monitored person, such as standing, walking, falling, and other behaviors, through the animation without seeing any identifiable image information.

[0040] In the entire workflow, the privacy computing visual sensing unit does not generate, store or transmit any form of complete image frame. All intermediate data are structured, non-image metadata that cannot be restored to an image with identifiable individual identity, thus achieving privacy protection at the data collection source.

[0041] To enhance the flexibility and applicability of the system, several technical variants are also provided in this embodiment. The edge computing module can use an ASIC dedicated chip instead of the combination of FPGA and MCU to further reduce power consumption and cost, suitable for large-scale deployment. The configuration of visual features can also be dynamically adjusted, for example, all body parts use white lines for display, but the semantic information of the parts is reserved for end users to customize rendering. The data output module can integrate PoE (Power over Ethernet) function to realize power supply and data transmission simultaneously through a single network cable, simplifying installation and wiring. In addition, the system can adjust the event threshold, clustering parameters and color scheme according to different application scenarios (such as home care for the elderly, medical rehabilitation) to adapt to diverse usage needs.

[0042] Through the above structure and processing flow, a visual sensing system with high privacy, low bandwidth and localized processing is realized, which is suitable for scenes with high privacy requirements such as home care and long-term care.

[0043] In one possible implementation, a certain nursing home deploys the privacy computing visual sensing unit system of the present application to improve the efficiency and safety level of daily care for the elderly residents. The nursing home has installed regular high-definition cameras in public areas such as corridors, dining rooms and activity rooms, achieving video monitoring of shared spaces. However, due to the protection of the privacy of the elderly, traditional cameras are not installed in the interior of the rooms where the elderly live and in private spaces such as bathrooms, forming a blind area for behavior monitoring. This results in the need for care workers to confirm the safety status of the elderly by regularly patrolling the rooms, usually entering the room for manual inspection every 15 minutes to determine whether there is a fall, long-term inactivity or other abnormal situations.

[0044] Frequent room patrols not only increase the workload of care workers, but also easily disturb the rest and daily private activities of the elderly, affecting their quality of life. To solve this problem, the nursing home introduces the privacy computing visual sensing unit system of the present application, installing the device on the top of each resident's room and bathroom to achieve non-imagery and non-intrusive monitoring of private spaces.

[0045] In the nursing station of the nursing home, a display terminal 103 is configured to run a special application program that receives and analyzes the visualization data stream 102 output from each room sensing unit in real time. The application program dynamically renders the data into desensitized humanoid animations, displaying the behavior status of the elderly in the form of abstract lines and geometric shapes. For example, the head is represented by a red dot, the upper limbs are represented by blue thin lines, the lower limbs are represented by green thick lines, and the torso is represented by a yellow outline frame. When the elderly are walking normally, the system displays a continuous motion trail; when a sudden drop or long-term inactivity is detected, the system automatically triggers an alarm to prompt the care workers to respond in a timely manner.

[0046] The system supports multiple visualization modes, which can be flexibly switched according to actual needs. In the conventional mode, color labeling is used to facilitate quick identification of behavior details; in the night or low interference period, it can be switched to grayscale mode, all parts are displayed with white lines, reducing visual stimulation; it can also be set to display only the moving target, and the static object automatically fades out, focusing on the current activity.

[0047] Unlike traditional passive monitoring systems, the system has significant technical innovation in user interface management. When a sensor in a room detects that a patient has started to move, the system will immediately sort and prioritize the monitoring window of that room on the monitoring interface of the display terminal 103. Specifically, among the multiple picture windows arranged in a fixed grid, the window of the room where the active event is occurring will automatically pop up to the main view area and display its desensitization human animation in a larger size, while adding a dynamic flashing border or status identifier (such as "active" or "moving") to the window edge. If multiple rooms are active at the same time, the system will automatically sort the windows according to pre-set rules (such as activity intensity, duration, or historical risk level) to place the most attention-worthy objects in the front row.

[0048] This mechanism allows nursing staff to notice individuals whose behavior is changing first, greatly improving information acquisition efficiency and emergency response speed. For example, when an elderly person sits up from bed and tries to walk, the picture of the room where he is located will be immediately pushed to the main screen, even if the window was previously in a small window or hidden state. This event-driven active visualization management strategy changes the passive mode of traditional monitoring systems that rely on manual round patrols or simple alarm prompts, achieving a technological leap from "people looking for information" to "information looking for people".

[0049] In addition, the system also has behavior trend analysis capability. If a potential risk behavior such as frequent night waking, long standing still, or unstable gait occurs in an elderly person, the system can prioritize their monitoring window and mark it as a "focus object" before reaching the emergency alarm threshold, assisting nursing staff in early intervention.

[0050] Since the system does not generate, store, or transmit any image frames throughout the process, all data is structured metadata that cannot be restored to a picture that can identify individual identities, fundamentally protecting the privacy rights of the elderly. Based on the reliable monitoring capability of the system, the nursing home adjusts the routine room patrol from every 15 minutes to every 180 minutes, and only conducts manual room inspection when the system alarms or special needs arise. This adjustment significantly reduces the frequency of nursing staff entering and exiting the room, avoids disturbing the elderly during their rest, and significantly reduces the cost of human patrol, improving service efficiency and care quality.

[0051] The practical application of the project shows that the privacy computing visual sensing unit system of the application not only fills the blind area of traditional monitoring in private space, but also realizes efficient, accurate and low-interference visual management through an intelligent interface scheduling mechanism. The output visual data stream 102 is transmitted to the display terminal 103 via the data output module, and the whole process from the original event stream 101 collection, edge computing module processing to the final display is completely presented. The three module block diagrams shown in the figure clearly show the data flow direction of "original event stream 101 → visual data stream 102 → display terminal 103", and the arrow clearly indicates the processing path of each link.

[0052] The system is different from a simple video monitoring alternative, and the core innovation is that under the premise of not sacrificing privacy, through an event-driven dynamic interface management mechanism, active presentation and priority sorting of key behavior information are realized, and monitoring efficiency and response ability are significantly improved.

[0053] In one possible implementation, the privacy computing visual sensing unit serves as a data acquisition end and does not generate, store or transmit any form of complete image frame in the entire workflow. This feature is realized through the coordinated design of hardware architecture and data processing mechanism, which fundamentally eliminates the generation of image-level privacy information.

[0054] The sensing unit uses an event-type image sensor chip such as Prophesee Gen4 or equivalent devices, and its physical working mechanism is based on the asynchronous event response principle. Each pixel independently monitors the local light change and generates an event signal containing spatial coordinates, time stamp and brightness polarity only when the change exceeds the set threshold, without periodic full-frame scanning and without frame buffer structure. Therefore, the sensor hardware itself does not have the ability to generate image frames, and the output data is naturally a sparse spatio-temporal event stream that cannot be reconstructed into a visual picture.

[0055] In order to continuously verify the core privacy attribute of "no image generation" during system operation, the application integrates a lightweight data format self-checking module inside the sensing unit. This module analyzes the data packet structure to be sent in real time before data output, and identifies whether there are image frame features through a pattern matching algorithm. Specifically, the module detects whether there are fixed periodic time intervals, regular pixel grid arrangements, frame header and frame tail synchronization markers or image encoding feature codes in the data stream. Once a data pattern similar to an image frame is detected, the system will immediately trigger an alarm and interrupt data transmission, ensuring that abnormal behavior is not transmitted externally.

[0056] The self-checking mechanism is a built-in function of the system, does not require external device intervention, runs on the embedded microcontroller of the sensing unit, has low resource occupation, responds in real time, and can be automatically activated with large-scale deployment of the device. Compared with external verification methods such as monitoring signals using a logic analyzer, manually auditing firmware code, relying on network packet capture tools or external decoding tests, this method does not require additional hardware support, does not increase significant operation and maintenance costs, and can achieve automatic privacy compliance verification.

[0057] Among them, the external logic analyzer and network packet capture method can effectively identify the data type, but it needs to be operated by professional personnel and is difficult to be continuously deployed in practical application. The firmware code audit belongs to static review and cannot cover the runtime behavior. The standard image decoder test is simple, but it is a post-verification method and cannot prevent potential risks. Therefore, the self-checking method not only serves as part of the system function, but also serves as a key technical means to realize "verifiable no image generation", thereby enhancing the credibility and compliance transparency of the system. Users or supervisors can confirm that the system is always in a no-image output state by calling the self-checking log, thereby establishing technical trust in the privacy protection mechanism.

[0058] As shown in Figure 2 The structure diagram of the original event stream data 202 output by the event type visual sensor 201 in the embodiment of the application is used to illustrate that each event contains a four-tuple data format of X coordinate, Y coordinate, timestamp and brightness change polarity.

[0059] The event type visual sensor is used to capture light changes in the environment. It does not shoot pictures at a fixed frame rate, but only records information when the brightness of a certain pixel point changes significantly. This working mode is completely different from the traditional camera, which continuously generates a whole image, while the sensor only outputs the changed part, so the data volume is smaller and the response speed is faster.

[0060] Whenever a pixel detects a brightness change exceeding a preset sensitivity threshold, a record is generated, which is called an event. Each event contains several key information. First, the position information, including the horizontal position and vertical position of the pixel in the image area, represented by X coordinate and Y coordinate respectively. The two coordinates each occupy ten binary data, which can accurately identify any pixel point within the range of one thousand twenty-four by one thousand twenty-four.

[0061] Next is the direction of the brightness change, which indicates whether the pixel is getting brighter or darker. This information is represented by one binary data, called polarity. If the brightness increases, it is marked as positive, and if the brightness decreases, it is marked as negative. In addition, the system reserves six data spaces for possible future functional expansion or data verification, to ensure the reliability of transmission.

[0062] Each event also contains a high-precision time stamp to record the exact time when the change occurs. The time stamp occupies thirty-seven bits of binary data, which can achieve a time resolution of microseconds or even higher, ensuring that all events can be correctly arranged and processed in chronological order.

[0063] These pieces of information are combined to form a structured data packet, the structure of which is shown in the following table:

[0064] | X (10bit) | Y (10bit) | Polarity (1bit) | Reserved (6bit) |Timestamp (37bit) |

[0065] The entire process is completed by the internal circuit of the sensor. When a pixel changes, its position, change direction, and occurrence time are quickly encoded and output. Pixels that do not change do not generate any data. Only when there is movement or light change in the environment, the system will generate data, and static scenes do not generate information flow.

[0066] The relationship between the sensor and the raw event stream is direct and efficient. The sensor 201 is responsible for sensing light changes and generating corresponding event data, and the raw event stream data 202 is an ordered set of these events. Each event truly reflects the brightness change at a certain time and location. Since no complete image frame is generated, these data cannot be restored to recognizable face or scene images, thus achieving privacy protection at the data source. This design is suitable for occasions that require long-term monitoring but must protect personal privacy, such as elderly care, home security, and other scenarios.

[0067] In one possible implementation, the raw event stream data collected by the event-based visual sensor is transmitted to a processing device with computing capability for analyzing the behavior state of the monitored target object. These processing devices can be local edge computing devices, remote cloud servers, or even ordinary personal computers or workstations, as long as they have the necessary computing resources and software environment to run artificial intelligence models.

[0068] The event-based visual sensor works in an asynchronous manner, generating data records only when the light intensity of a pixel point changes significantly, and each record contains the location of the event, the direction of the brightness change, and an accurate time stamp. These data do not constitute traditional image frames, but exist in the form of sparse, timestamp-based point events. Multiple events are combined in chronological order to form a raw event stream. Since the system does not generate complete pictures from the source, it is impossible to restore visual information that can identify individual identities, which provides a foundation for privacy protection for subsequent data processing.

[0069] The data output by the sensors is encapsulated in a structured format for ease of parsing and processing by subsequent systems. One of the supported data formats is JSON. JSON is a lightweight, easy-to-read and transmit text format widely used for data exchange between different systems. By encapsulating each event or batch of events as a JSON object, the system can clearly express fields such as the X coordinate, Y coordinate, polarity, timestamp, etc. of the event, while also supporting the addition of metadata such as device number, time zone, data version, etc. This structured output method allows the data to be directly read and processed by general software systems, databases, or artificial intelligence models, without relying on specialized image decoders or video players.

[0070] Although the sensor primarily operates in event mode and does not generate continuous image frames, the system may also support the output of traditional video format as a supplementary function in certain specific application scenarios. For example, during the debugging phase or when compatibility with other systems is required, the device can provide a video stream with a resolution of 1280 by 720 and 20 frames per second. This video mode is not intended for daily monitoring, but rather to assist in system calibration or verification of the accuracy of event data when necessary. Even when working in video mode, the system can choose to generate video segments only when motion is detected, thereby reducing data volume and storage pressure. Importantly, event mode and video mode can be independently switched or run in parallel, allowing users to flexibly configure according to privacy requirements and usage needs.

[0071] After these event data are sent to the processing device, they first undergo a preprocessing stage. The system aggregates events within a certain time window (e.g., every 50 milliseconds) in space and time to generate a two-dimensional representation called an "event frame." In the event frame, each pixel position records the number of events that occurred in that area within the time window and their polarity (brightening or darkening). This conversion makes the originally sparse and asynchronous data structured, allowing it to be effectively processed by subsequent artificial intelligence models.

[0072] In the processing device, a convolutional neural network is used to analyze the event frames. Convolutional neural networks are machine learning models that are good at recognizing spatial patterns, and they gradually extract local features such as motion direction, speed change, and spatial distribution patterns through multiple layers of neurons. Through training, the model can learn the event distribution patterns of specific behaviors. For example, when a person suddenly falls off a bed, a high-density area moving rapidly downward will appear in the event frame, and the model can recognize this behavior by comparing it with known bed-falling samples and trigger an alarm. Similarly, when there is rapid, intense, and directionally chaotic event activity between two targets, the system can judge it as a fight. For the case of entering a prohibited area, the system can mark the prohibited entry range in the spatial coordinates, and when it detects that event activity continues to occur within this area, it considers it as a violation of entry and issues a reminder.

[0073] To make the model capable of recognition, training is needed. The training process uses a large amount of labeled event data from real or simulated use scenarios, and the corresponding behavior categories such as walking, falling, falling out of bed, fighting, etc. are labeled by professionals. The model continuously learns on these data, adjusts internal parameters, until it can accurately distinguish different behavior patterns. Training can be done on high-performance computing platforms or in the cloud, and after completion, the model can be deployed on various devices for inference.

[0074] In practical applications, the deployment of the model has high flexibility. For example, in one implementation case, the system uses an embedded chip module integrated with a dedicated neural network acceleration unit, which has the characteristics of low power consumption and high real-time performance, suitable for local data processing to avoid data transmission and further enhance privacy security. In another implementation case, the processing task is handed over to a local personal computer or industrial computer, which uses its strong general computing power to run the analysis model. In addition, the system also supports uploading pre-processed structured data to a remote server or cloud platform for further analysis by more complex deep learning models. For example, through long-term accumulation of behavior data, the system can learn the individual's daily activity rules, identify abnormal deviations, and achieve higher-level health monitoring and risk warning.

[0075] The design of the entire system does not depend on any specific hardware model, but is based on a general data processing process and an artificial intelligence analysis framework. Whether using embedded devices, ordinary computers or cloud servers, as long as they can receive raw event streams, generate event frames and run trained models, the behavior recognition function can be realized. The flexibility of this architecture makes the technology adaptable to various scenarios from small-scale deployment at home to large-scale monitoring in institutions, with wide applicability and scalability.

[0076] During deployment, the system also needs to be thoroughly tested to ensure stable operation under different environmental conditions (such as changes in light, multiple activities, partial occlusion, etc.). All data transmission uses encryption mechanisms to prevent information leakage. By combining event-based sensor technology with artificial intelligence analysis, this embodiment realizes efficient and accurate identification of abnormal behavior while protecting individual privacy.

[0077] In one embodiment, the visual sensing unit described in the present invention is deployed in an artificial intelligence research project aimed at understanding the daily behavior of the elderly, aiming to collect real, natural and continuous behavior data while strictly protecting individual privacy, and generate a structured desensitized behavior label set for machine learning model training.

[0078] Multiple privacy-compliant visual sensing units are installed in an experimental site simulating a home environment, including living room, bedroom, kitchen, and bathroom. Volunteers enter the scene after fully understanding the research purpose and signing a written informed consent form, and simulate daily life activities freely. Each sensing unit works based on event-based vision principle, only responds to changes in scene brightness and outputs an asynchronous event stream. This event stream is processed in real time by an edge computing module locally, and is converted into non-imaging behavior representation data, without generating any image frames or video streams.

[0079] The generated behavior data includes the two-dimensional motion trajectory of the target object, body posture estimation results (such as standing, bending, squatting), start and end time markers of key actions, and semantic behavior labels. The behavior labels cover twelve common daily activities, including standing, sitting, bending, squatting, lying, walking, drinking, eating, using a mobile phone, opening a door, picking up an object, and turning around. These labels are defined according to the international standard for human activity classification, ensuring semantic consistency and cross-system comparability.

[0080] All behavior data is organized and stored in a structured format, serialized using the lightweight data exchange format JSON. Each record contains the following fields: timestamp (accurate to milliseconds), anonymized subject number (subject_id), behavior category label (activity_label), action confidence (confidence_score, value range 0.0 to 1.0), duration (duration, unit: seconds), spatial position coordinates (x, y), and data source device ID (device_id). The data format design supports efficient retrieval, batch processing, and long-term archiving.

[0081] After preliminary cleaning and alignment of the data in the local edge server, the data is checked for compliance by the image frame leakage protection device described in the invention, ensuring that the output stream does not contain any image features or identifiable individual identity information. Data that passes the compliance check is encrypted and stored in the local database, and can be uploaded to the centralized research platform through a secure channel.

[0082] The collected desensitized behavior data supports two artificial intelligence model training modes. The first is the local training mode, which uses the data output by the local image frame leakage protection device to train a lightweight behavior recognition model on the edge computing node deployed in the experimental site, for real-time fall detection, abnormal behavior warning, and other local applications. The second is the cloud aggregation training mode, which uploads the desensitized data to the centralized research platform after obtaining the explicit authorization of the volunteers, aggregates the data with other experimental sites, and trains a general behavior understanding AI model that can be used across scenes and populations, improving the model's generalization ability.

[0083] The embodiment has completed a pilot deployment in an old-age health research base, and has accumulated more than 200 hours of effective behavior data, covering a volunteer group of different ages, genders and action abilities. Experimental results show that the AI model trained based on the desensitization data collected by the system in the application has an accuracy of 92.3% in the fall detection task, and does not need to rely on manual video annotation, significantly reducing the data annotation cost and privacy leakage risk.

[0084] All data storage complies with the principles of retrievability, traceability and auditability, and supports efficient queries by time range, behavior type, subject number and other dimensions. The data retention period is specified by the research ethics committee, and is automatically cleared after expiration. The embodiment provides a safe, compliant and sustainable technical path for data collection and model training in privacy-sensitive scenarios.

[0085] Figure 3 The original event stream and frame image data in the embodiment of the application are compared in the motion capture effect, which aims to illustrate the advantages of event data with microsecond time resolution and no motion blur.

[0086] When the traditional frame camera captures the fast-moving ball 301, when the ball moves at high speed, due to the limitations of the exposure time and sampling frequency of the frame camera, the image captured has obvious blurred trailing. This phenomenon makes it difficult to accurately capture the motion trajectory of the ball, affecting the accuracy and reliability of subsequent analysis. The trajectory left by the ball in the picture shows continuous and unclear lines, because each frame of image contains all the position information of the object in this period of time, not a single accurate position point.

[0087] In contrast, in the same scene, the event sensor outputs a sparse point sequence 302. When the ball moves at the same speed, the event sensor can capture every tiny change of the ball with extremely high time resolution. These sparse point sequences accurately restore the motion trajectory of the ball without blur or trailing phenomenon. Each point represents the exact position of the ball at a specific time, which can clearly depict the actual motion path of the ball. This high-precision data capture method is particularly suitable for application scenarios that require accurate motion tracking.

[0088] From the comparison, it can be seen that although the traditional frame camera can also capture the general motion direction of the object, it is far inferior to the event sensor in detail capture. The event sensor not only records the timestamp of the event occurrence on the microsecond time scale, but also provides more accurate and blur-free motion trajectory capture capability, providing high-quality data support for behavior recognition and dynamic scene analysis. This feature makes the event sensor have a significant advantage in handling high-speed motion scenes, especially suitable for applications that require high precision and real-time performance.

[0089] In one possible implementation, an insurance company provides a value-added service to users who purchase accidental insurance. In order to reduce the risk of users falling or other accidents in the home environment, the insurance company provides a set of privacy computing-based visual sensing unit behavior monitoring systems to users for free. The system is installed in key areas such as the bedroom, kitchen, living room, and bathroom of the user's home to perceive the activity state of the occupant in real time and issue a timely warning when an abnormal behavior is detected.

[0090] The core component of this system is an event-based visual sensor, which does not capture images or videos in the traditional way, but only records data when the ambient light changes. Whenever the brightness of a certain pixel point changes significantly, such as someone moving, an object falling, or light and shadow changing, the sensor generates a data point with location and time information, called an "event". These events are arranged in chronological order to form an original event stream, which is then processed and analyzed by a local edge device. The entire process does not generate any complete picture, so it is impossible to restore identifiable facial or body features, fundamentally avoiding the problem of infringing personal privacy.

[0091] In actual deployment, the system can use Sony's EVS series of event-based visual sensors, such as IMX636 or its subsequent models. These sensors are commercial products developed by Sony based on its proprietary event-driven technology, with characteristics such as high temporal resolution, wide dynamic range, and low latency response. According to the published technical information, Sony's EVS sensor uses a pixel-level asynchronous readout mechanism, where each pixel independently determines whether the brightness change exceeds a threshold, and once triggered, it outputs event data with a microsecond-level timestamp. This design allows the sensor to work stably in extremely dark or strong light contrast environments, making it particularly suitable for complex lighting scenarios in the home, such as getting up at night or changes in window curtain light.

[0092] Compared with traditional high-definition cameras, this type of sensor has an essential difference. High-definition cameras continuously capture the entire image at a fixed frame rate, generating dozens of frames per second. Whether there is activity or not, it will continue to record. This means that users' private behaviors, such as dressing, resting, or washing, will be completely captured and stored. Even if the data is encrypted, as long as the image exists, there is always a risk of being restored or leaked. Many users therefore refuse to install traditional surveillance, making it difficult for insurance companies to obtain real behavior data for risk assessment and claim underwriting.

[0093] The Event-Based Vision Sensor (EVS) used in this system is an innovative image sensing technology that monitors brightness changes in real-time through a pixel array. When the brightness change exceeds a set threshold, the sensor outputs event data. It has low latency, high dynamic range, and low power consumption. Its output is only a sparse event data stream, containing no image frames. This data expresses spatial position, brightness change direction, and occurrence time in a structured manner and cannot be restored to a visually readable picture. Even if the data is intercepted, individual identities or specific behavior details cannot be identified, truly achieving "imageless monitoring." This technical approach not only meets the insurance company's demand for accurate capture of risk events but also fully respects users' privacy rights, significantly improving user acceptance and system feasibility.

[0094] When a user uses a stool at home and stands on a high place to get something, the system can perceive the whole process of this behavior. For example, when the user moves the stool, steps on the stool surface, and stretches up his hand, the event sensor will record a series of continuous position changes and motion trajectories. If the user loses balance and falls suddenly during this process, the system will detect a rapid downward event cluster in a very short time, followed by a long period of inactivity. After real-time analysis by the local edge device, this data is confirmed as a high-risk fall event and triggers a local alarm immediately, while uploading the desensitized structured data to the insurance company's backend system.

[0095] Another common situation is that the user quickly walks or runs at home, especially when the ground is slippery or there are obstacles, which can easily cause a fall due to unstable center of gravity. Traditional cameras may miss critical moments due to frame rate limitations or have unclear trajectories due to motion blur. The event sensor captures every brightness change with a time resolution of microseconds, accurately restoring the entire process from normal walking to sudden imbalance, body tilt, and finally falling to the ground. The artificial intelligence model can accurately determine whether this is an accidental fall or normal squatting or sitting by analyzing the time density, spatial distribution, and motion direction changes of events.

[0096] These analyzed behavior data become important evidence for insurance companies to review claims. The information provided by the system includes the specific time, location, duration, and action characteristics of the event, all of which are structured metadata and do not involve any image content. Insurance companies can combine these objective data to determine whether the accident falls within the scope of insurance liability, improve underwriting efficiency, and reduce disputes. At the same time, long-term accumulation of behavior data can also be used to evaluate users' living habits and risk levels, providing support for the design of personalized insurance products.

[0097] The application scenarios of this technology are not limited to home insurance services, but can also be widely used in nursing homes, rehabilitation centers, monitoring of elderly people living alone, smart communities, and other occasions that are sensitive to privacy but require security guarantees. The system supports multiple deployment methods, and can complete all calculations on local edge devices or upload structured results to the cloud for centralized analysis. The edge device can be equipped with a general-purpose computing platform, such as an embedded module with a neural network acceleration unit, or a general industrial computer or personal computer, with good compatibility and expandability.

[0098] The design of the entire system is not limited to a specific chip, and the event-based visual sensor is one of the preferred solutions in this embodiment. This privacy computing visual sensing system ensures user safety while ensuring that "danger can be seen, but privacy cannot be seen." For insurance companies, this not only improves the accuracy of underwriting and claims settlement, reduces fraud risk, but also enhances customer trust and service value.

[0099] In one possible implementation, in order to reduce the risk of users falling or other accidents in a home environment, an insurance company provides a behavior monitoring system based on a privacy computing visual sensing unit to the user. The system is installed in key areas such as the bedroom, kitchen, living room, and bathroom of the user's home, and is used to perceive the activity state of the occupant in real time and issue a warning in a timely manner when an abnormal behavior is detected.

[0100] The core component of this system is an event-based visual sensor, which does not capture images or videos in the traditional way, but only records data when the ambient light changes. Whenever the brightness of a certain pixel changes significantly, such as when someone moves, an object falls, or the light changes, the sensor generates a data point with location and time information, called an "event". These events form a raw event stream in chronological order, which is then processed and analyzed by a local edge device. The entire process does not generate any complete picture, so it is impossible to restore identifiable facial or body features, fundamentally avoiding the problem of infringing on personal privacy.

[0101] In actual deployment, the system can use a visual sensor chip with event-driven features, for example, the IMX636 produced by Sony is a CMOS image sensor specifically designed for event-based vision, with a pixel array of 1280 by 720 and a microsecond-level time resolution, capable of responding to brightness changes with extremely low latency. The chip uses an asynchronous pixel readout mechanism, where each pixel works independently and only outputs event data when it detects a brightness change that exceeds a threshold, significantly reducing power consumption and data volume. This sensor has a high dynamic range and can operate stably in environments with dramatic changes in light, making it suitable for a variety of scenarios in the home, from bright living rooms to dim hallways.

[0102] Besides IMX636, other sensor chips of the same series or similar architecture can also be used in the system as long as they support event stream output mode and have sufficient time accuracy and spatial resolution. The common feature of these chips is the ability to output structured data packets containing X coordinate, Y coordinate, polarity (brightness change direction) and high-precision timestamp, without generating traditional image frames. This form of data output provides a basis for subsequent privacy protection and intelligent analysis.

[0103] Compared with traditional high-definition cameras, such sensors have essential differences. High-definition cameras continuously capture the entire image at a fixed frame rate, generating dozens of frames per second, regardless of whether there is any activity. This means that users' private behaviors, such as dressing, resting or washing, will be completely recorded and stored. Even if the data is encrypted, as long as the image exists, there is always a risk of being restored or leaked. Many users therefore refuse to install traditional monitoring, making it difficult for insurance companies to obtain real behavior data for risk assessment and claim underwriting. The technical path adopted by the system outputs only sparse event data streams, without any image frames. These data express spatial position, brightness change direction and occurrence time in a structured manner, and cannot be restored to visually readable pictures. Even if the data is intercepted, it is impossible to identify individual identity or specific behavior details, truly realizing "imageless monitoring". This design not only meets the needs of insurance companies for accurate capture of risk events, but also fully respects users' privacy rights.

[0104] When the user uses a stool and stands at a high place to get something at home, the system can perceive the whole process of the behavior. For example, when the user moves the stool, steps on the stool surface and stretches his hand upwards, the sensor will record a series of continuous position changes and motion trajectories. If the user loses balance and suddenly falls during this process, the system will detect a rapid downward event cluster in a very short time, followed by a long period of inactivity. After real-time analysis by the local edge device, these data are confirmed as a high-risk fall event, and a local alarm is triggered immediately, while the desensitized structured data is uploaded to the insurance company's backend system.

[0105] Another common situation is that the user quickly walks or runs at home, especially in the case of slippery ground or obstacles, which is easy to fall due to unstable center of gravity. Traditional cameras may miss the key moment due to frame rate limitations, or the trajectory may not be clear due to motion blur. Event sensors capture every brightness change with microsecond-level time resolution, accurately restoring the whole process from normal walking to sudden imbalance, body tilt and finally falling. Artificial intelligence models can accurately determine whether it is an accidental fall or normal squatting or sitting by analyzing the time density, spatial distribution and motion direction change of events.

[0106] These analyzed behavior data become an important basis for insurance companies to review claims. The information provided by the system includes the specific time, location, duration, and action characteristics of the event, all of which are structured metadata and do not involve any image content. Insurance companies can combine these objective data to determine whether the accident falls within the scope of insurance coverage, improve underwriting efficiency, and reduce disputes. At the same time, the long-term accumulation of behavior data can also be used to evaluate the user's living habits and risk level, providing support for the design of personalized insurance products.

[0107] The application scenarios of this technology are not limited to home insurance services, but can also be widely used in nursing homes, rehabilitation centers, monitoring of elderly people living alone, smart communities, and other occasions that are sensitive to privacy but require security. The system supports multiple deployment methods, and can complete all calculations on local edge devices or upload structured results to the cloud for centralized analysis. Edge devices can be equipped with general computing platforms, such as embedded modules with neural network acceleration units, or ordinary industrial computers and personal computers, with good compatibility and expandability.

[0108] In some specific implementation cases, the core innovation of the present application is to build a privacy-first behavior monitoring system based on event-type sensors. This system realizes accurate identification of high-risk behaviors through non-image data collection methods combined with edge intelligent analysis technology. In other implementation cases, the core innovation is reflected in the design of the data processing flow, including the conversion of asynchronous event streams into event frames that can be processed by artificial intelligence models, and the use of convolutional neural networks to classify behavior patterns. In addition, in specific deployment scenarios, the innovation of the system is also reflected in the deep integration with insurance business processes, such as using structured event data as objective evidence for claim review and underwriting, thereby improving service efficiency and risk management capabilities.

[0109] Figure 4 The processing flow of the edge computing module in the embodiment of the present application for visual identification of human body parts is shown, which is used to illustrate the complete technical path from the original event stream to the generation of the desensitized graphics. It embodies how the system realizes the identification and expression of human motion structure through intelligent analysis without relying on traditional images.

[0110] The event-type visual sensor continuously outputs the original event stream, and each event contains the location, direction of brightness change, and accurate time marker. These events are generated in an asynchronous manner, forming an event stream 4011. Since the data is sparse and scattered, the system first performs spatial clustering processing on these events, i.e., spatial clustering 4012. This process groups events that are close in time and adjacent in space into a group through an algorithm, identifying the area where activities are concentrated in a short period of time. This clustering method can effectively filter out noise and highlight the local high-density event clusters formed by human motion.

[0111] After spatial clustering, the system extracts multiple active regions 4013. These regions represent the dynamic activity range of different parts of the user's body in space. For example, rapid hand movement forms an elongated event-dense band in the image region, while slow torso movement appears as a larger area with lower event density. The identification of active regions provides a basis for subsequent analysis of human structure.

[0112] Next, the system enters the motion consistency analysis phase, i.e., motion consistency analysis 4021. In this process, the algorithm extracts dynamic features from the event sequence in each active region, analyzes its motion direction, speed variation, and time continuity. By comparing the motion patterns between different regions, the system determines which regions are in coordinated motion. For example, the swing of the arm usually maintains consistent motion direction and rhythm with the shoulder region, while the slight movement of the head exhibits independent trajectory characteristics.

[0113] Based on the results of motion consistency analysis, the system divides the human structure, i.e., divides the head, upper limbs, lower limbs 4022. This process relies on a pre-trained machine learning model that learns a large number of event distribution patterns of human motion, and can map active regions to specific body parts according to motion characteristics. For example, a region with fast, small amplitude, and periodic motion may be identified as a hand, while a region located above, relatively stable but with slight shaking, may be determined as a head. This division does not depend on image contours or skin color information, but is completely based on the spatiotemporal characteristics of dynamic events, thus having strong robustness and privacy security.

[0114] After completing the division of human body parts, the system enters the semantic mapping phase. At this time, each identified body part is assigned specific visualization attributes, such as color or line width, i.e., color / line width allocation 4031. For example, the head can be represented by a thick red line, the upper limbs by a medium-width blue line, and the lower limbs by a green line. These attributes are only used to distinguish the motion state of different parts and do not contain any texture or detail information.

[0115] Finally, the system generates a desensitized graph 4032. This graph consists of several lines or skeleton structures, reflecting only the relative positions and motion trajectories of different parts of the human body. The entire graph does not contain pixel-level image information and cannot be restored to identifiable facial or body features, thus belonging to a highly desensitized expression form. This graph can be used for local display warnings and can also be uploaded to the server for behavior analysis.

[0116] The entire processing flow is completed on the edge computing module, without the need to transmit raw data out, further enhancing the security and real-time performance of the system. This technical path not only realizes accurate perception of user behavior, but also avoids privacy risks brought by traditional cameras through structured expression.

[0117] In one possible implementation, the present application provides an integrated privacy computing visual sensing device, which adopts an integrated metal or engineering plastic shell design, and highly integrates an event-type visual sensor chip and a processing module with artificial intelligence computing capability inside. The device has a compact overall structure, is suitable for wall-mounted or ceiling-mounted deployment, and can realize unobtrusive behavior monitoring in typical living scenes such as bedrooms, bathrooms, and corridors. The core perception unit adopts a Sony EVS series event-type visual sensor chip, such as IMX636, which has a 1280 by 720 effective pixel array, supports asynchronous event output, has a time resolution of microseconds, a dynamic range of more than 120 dB, and can work stably in extremely low light or strong light contrast environments, ensuring accurate capture of human activity in night or alternating light and dark areas.

[0118] Inside the device, directly connected to the Sony IMX636 sensor is an AI processing chip with neural network acceleration capability, which can be RK3588 of Rockchip in this embodiment. The chip adopts an 8-nanometer process, has a four-core Cortex-A76 and four-core Cortex-A55 CPU architecture, integrates a Mali-G610 GPU, and is equipped with a dedicated 6TOPS computing power NPU (neural network processing unit), which can efficiently run complex models such as convolutional neural networks and spatiotemporal clustering algorithms. The raw event stream output by the sensor is directly transmitted to the RK3588 chip through the MIPI interface, and all data processing is completed locally without relying on external networks or cloud servers, realizing true edge intelligence.

[0119] The working process of the device strictly follows the technical path shown in Figure 4 The raw event stream is first sent to the NPU unit of RK3588 for spatial clustering processing. The system identifies the areas where events occur densely in a short time through algorithm, forming multiple active areas. These areas correspond to the motion trajectories of different parts of the human body, such as arm swinging, leg moving, or torso displacement. Then, the system uses motion consistency analysis technology to extract dynamic features of the event sequence of each active area, including motion direction, speed change, and time continuity, to determine which areas are in coordinated motion.

[0120] Based on the above analysis, the system divides the human structure and maps the event cluster to semantic parts such as head, upper limbs and lower limbs. This process relies on a pre-trained lightweight deep learning model, which has been trained on a large amount of event data in real home scenes and can accurately identify typical behavior patterns. After completing the part division, the system assigns specific visualization attributes such as color or line width to each part, and finally generates a desensitized graph containing only structured motion information. This graph is presented in the form of skeleton lines or dynamic trajectories, and does not contain any pixel-level image information, so it cannot restore identifiable facial or body features, ensuring user privacy.

[0121] In another possible implementation, the Sony IMX636 sensor chip is not integrated with the AI processing unit in the same physical housing, but transmits the raw event stream to an external edge computing device for centralized processing through a high-speed interface. This edge computing device can be a home local server, NAS storage device or dedicated smart gateway, with strong computing power and large storage space. In this device, it also carries a RIKKAI micro RK3588 or an AI chip with similar performance, runs the same analysis model as the integrated device, and completes the Figure 4 processing flow shown.

[0122] The edge computing device also has data storage and visual playback functions. When the system detects abnormal events such as falls or long periods of inactivity, the relevant desensitized graph, event timestamp and behavior label are encrypted and stored in the local hard disk. Users or authorized personnel can access these records through a matching mobile application or Web interface to view the time, location and action evolution of the event. Since the storage content is only structured metadata and desensitized graphs, it does not contain any image frames, so even if the device is physically accessed, the user's private visual information cannot be obtained.

[0123] The overall system architecture design takes into account privacy protection and functional integrity. Whether it is an integrated integrated solution or a distributed edge deployment, it adheres to the principle of non-image perception, avoiding the acquisition of any form of visual picture. At the same time, the system is compatible with a variety of sensors that meet the event output specification through a standardized interface, not limited to a specific manufacturer or model, enhancing the universality and scalability of the technology.

[0124] Figure 5 A frame structure diagram of the privacy computing visualization data stream output in the embodiment of the application is shown, which details the lightweight binary format 501 containing target number, position, motion vector and part feature array. This design aims to transmit the most necessary information with the smallest amount of data, while ensuring efficiency and reliability during storage and network transmission.

[0125] First, let's look at the frame header 5011 part. The frame header is represented by 8 bits, which is used to identify the start of the data packet and carries some basic control information such as data type or check code. This helps the receiving end quickly identify the starting position of the data packet, ensuring the accuracy of data parsing.

[0126] Next is the timestamp 5012 part, which uses 37 bits to record the specific time of the event. High-precision timestamps are crucial for behavior analysis, especially when tracking continuous action sequences or performing multi-sensor data fusion. With time markers accurate to the microsecond level, the system can accurately restore the timeline of events, providing a reliable basis for subsequent behavior pattern recognition.

[0127] The target number 5013 part is represented by 8 bits, indicating the number of targets detected in the current frame. This field helps the receiving end pre-allocate enough memory space, avoiding buffer overflow problems caused by changes in the number of targets. In practical applications, the number of targets usually does not exceed 255, so 8 bits are sufficient to meet the demand.

[0128] The information of each target is encapsulated in the target 5014 part, which includes ID, type, X coordinate, Y coordinate, horizontal velocity Vx, vertical velocity Vy, and part array. Among them, the ID field uses 16 bits to represent, ensuring that each target has a unique identifier throughout the monitoring period, making it easy to track the motion trajectory of the same target between different frames. The type field uses 4 bits of encoding to distinguish between human bodies, objects, or other specific categories, simplifying subsequent processing logic.

[0129] Position information is composed of X and Y coordinates, each occupying 10 bits, representing the horizontal and vertical coordinate values of the target in a two-dimensional plane. This design balances precision and data volume, providing sufficient positioning accuracy in most indoor scenarios. Motion vectors Vx and Vy are each represented by 16 bits, describing the instantaneous speed of the target in the horizontal and vertical directions, supporting real-time analysis of dynamic behavior.

[0130] The part array is one of the most critical parts of the target 5014, which contains feature information of human body parts such as head, hand, and leg. For example, the head is represented by a red circle, the hand is represented by a blue line, and the leg is represented by a green thick line. These features not only intuitively reflect the body posture of the target, but also provide rich input data for advanced behavior recognition algorithms. The specific implementation of the part array may involve multiple subfields, each corresponding to the attributes of a part, such as color, shape, and relative position.

[0131] In one possible implementation, the visual sensing system provided by this invention is applied in medical research institutions for long-term observation and data analysis of the home behavior of specific populations. The system is deployed in the living environments of Alzheimer's patients or individuals undergoing drug withdrawal treatment to non-invasively collect their daily activity patterns, providing objective and continuous data support for researchers studying disease progression, behavioral abnormalities, and rehabilitation outcomes.

[0132] In practical applications, research institutions install event-based visual sensor-based monitoring devices in key areas of patients' residences, such as living rooms, bedrooms, and hallways. These sensors utilize Sony's EVS series chips, such as the IMX636, capable of capturing changes in ambient brightness with microsecond-level temporal resolution and outputting sparse event stream data. Unlike traditional high-definition cameras that continuously capture video at 30 frames per second, event sensors generate data only when motion occurs, resulting in extremely small data volumes. For example, a 12-hour home activity might generate over 100GB of video files with a traditional 1080p camera, while the structured event data generated by this system is typically less than 100MB, saving over 99% of storage space.

[0133] These raw event data are sent to local edge computing devices, which are equipped with AI chips such as the Rockchip RK3588 that have neural network processing capabilities, to execute... Figure 4 The processing flow is shown below. The system first performs spatial clustering on the event stream to identify active regions. Then, it uses motion consistency analysis to segment parts such as the head, upper limbs, and lower limbs, ultimately generating anonymized skeletal graphics. These graphics do not contain any identifiable facial or body details, fundamentally avoiding the risk of privacy leaks. In contrast, even after blurring or pixelating, videos recorded by high-definition cameras can still reveal identity information through image enhancement techniques, posing serious ethical and legal risks.

[0134] After processing, the system follows Figure 5 The frame structure shown encapsulates behavioral data into a lightweight binary format for storage. Each data packet contains a timestamp, target number, location coordinates, motion vector, and an array of part features. The overall structure is compact, with a single frame typically less than 100 bytes. This efficient data encoding method enables long-term continuous recording. Research institutions can store months or even years of behavioral data on local servers or private cloud platforms at a much lower storage cost than video surveillance systems. Due to the small data size, even when transmitted remotely over the internet, bandwidth consumption is low, making it suitable for stable operation in home broadband environments.

[0135] For Alzheimer's patients, researchers can identify typical symptoms such as memory decline, nighttime wandering, and repetitive actions by analyzing their long-term behavioral patterns. For example, the system can detect that a patient frequently gets up at night, walks back and forth on a fixed route, or spends a long time in the kitchen without completing cooking actions. These behavioral characteristics, when converted into structured data, facilitate statistical analysis and modeling by researchers to assess disease progression speed or the effectiveness of interventions.

[0136] For patients undergoing drug withdrawal, the system can monitor abnormal behaviors triggered by emotional fluctuations, such as pacing, sudden falls, or prolonged immobility. These data help healthcare professionals judge the severity of withdrawal reactions and adjust treatment plans in a timely manner. Since the entire process does not require continuous human observation, it not only reduces the nursing burden but also avoids the bias caused by subjective judgment.

[0137] All stored data can be retrieved and replayed by authorized researchers through a secure interface. The system provides visualization tools to restore historical data into dynamic desensitized graphic sequences, allowing researchers to visually view the behavior evolution process within a specific time period. Since the data itself does not have identifiable characteristics, research institutions do not need additional desensitization when collaborating across teams or publishing results, greatly simplifying the data sharing process.

[0138] Figure 6 To present the rendering effect of privacy computing visual data on the display terminal in the embodiments of the present application, we can clearly understand the difference between the privacy protection method of the present application and the traditional method by comparing Figure 601, Figure 602 and Figure 603. The original image 601 shows the result output by the traditional camera, which contains all the captured information, such as the facial features of the person, the style of the clothes, and the background details, etc. This direct visual record, although providing detailed scene information, has significant shortcomings in terms of personal privacy protection, as it may inadvertently reveal unnecessary personal information.

[0139] Next is the image 602 processed by the patented technology of the present application, which shows the result after special processing to protect user privacy. Unlike the original image 601, the personal features in this image have been removed or blurred, such as the head replaced by a highlighted white circle, the hands represented by thin gray lines, and the legs represented by thick gray lines. These changes not only help protect the identity of individuals from being identified, but also effectively demonstrate the motion state and behavior pattern, meeting the analysis needs while protecting personal privacy.

[0140] Finally, Figure 603 further demonstrates the application of the present invention, which focuses on recording the luminance changes within the scene rather than generating specific images. By capturing the changes in luminance and converting them into metadata (such as location, time, etc.), this approach avoids the generation of specific images, thus maximizing the protection of users' privacy. This method is particularly suitable for research and application scenarios that require a high level of privacy protection.

[0141] The original image 601 represents the direct output of a traditional camera, while images 602 and 603 are the results obtained after applying the present invention, aiming to protect personal privacy while providing effective data analysis support. This processing approach allows the present invention to strike a balance between privacy protection and the provision of useful data.

[0142] Figure 7 This figure demonstrates the visual output of privacy computation in different application scenarios, highlighting the system's ability to dynamically recognize standing, walking, and object retrieval behaviors within a room. The figure is divided into three columns, each representing a typical behavior pattern and expressed through simplified human silhouettes.

[0143] Standing 701: In this column, the human silhouette appears in a static state, with limbs hanging vertically on both sides of the torso. This representation clearly conveys that the individual is stationary, without relying on specific image details to understand the behavior of the person.

[0144] Walking 702: This column shows the characteristics of a person walking, with leg lines exhibiting a forward and backward swinging posture, with continuous and smooth trajectories. By observing these simple line movements, we can easily recognize that this is a person walking, without needing to see any specific facial or body features.

[0145] Bending to pick up 703: The last column depicts the individual's action of bending to pick up an object. The upper limbs are extended forward, the torso is noticeably tilted in a certain direction, and the head position is lowered to simulate the action of bending to pick up an object. Even without actual image information, such a simplified representation is sufficient for the observer to understand that the current action is picking up an object from the ground or a low place.

[0146] Figure 8 This figure is a deployment diagram for the sensor-to-cloud connection in the embodiment of the present invention. The sensor-to-cloud deployment scheme connects visual sensors 201 directly to the cloud, which capture the necessary visual information in a specific environment. These sensors are designed to recognize and track human silhouettes and movements without retaining or transmitting any information that can directly identify an individual's identity, ensuring privacy protection.

[0147] Once the visual sensor 201 finishes data collection, it transmits the collected data directly to the cloud server 801 using existing wireless communication technologies such as WiFi or 4G networks. This approach allows real-time data upload, reduces latency, and improves data processing efficiency.

[0148] After the cloud server 801 receives the raw data, it uses advanced algorithms to analyze, process, and de-identify the data. All specific features that can be used to identify individual identities are removed, leaving only abstract information such as behavior patterns. Subsequently, these processed data are converted into anonymous human animations for easy understanding and presentation. The processed anonymous human animations are pushed to the user's terminal device 802 in real time through the Internet, such as smartphones, tablets, or personal computers. Users can view dynamic behavior information through these devices without worrying about privacy leakage.

[0149] The entire process emphasizes the importance of privacy protection at every stage from data collection to final presentation to users. At the same time, this approach also demonstrates that modern technology can provide efficient and accurate service capabilities while protecting user privacy. This provides new solutions for smart home, security monitoring, and other fields, making people's lives more convenient and secure.

[0150] In one possible implementation, multiple event-type visual sensors are deployed in different positions in the same monitoring space, such as installed on the opposite walls or ceiling corners of the room, forming a multi-node collaborative perception network. Each sensor uses an event-driven visual chip, such as Prophesee Gen4 or iniVation DAVIS346, which can asynchronously respond to pixel-level light changes with microsecond-level time resolution, output only sparse time-space-polarity event streams, and does not perform full-frame scanning or collect static background information. Therefore, it is impossible to reconstruct any complete image that can identify individual identities, thus achieving privacy protection from the data collection source and meeting the requirements of global mainstream privacy laws for sensitive information processing.

[0151] These sensors collectively cover the target area for capturing human gesture behavior. For example, raising the arm above the head and waving up and down can be defined as triggering a help signal. Writing the number 110 in the air with the index finger continuously can be used as an emergency alarm instruction. And stretching the palm and then slowly clenching the fist can be used to turn off all alarm systems. Traditional frame-based camera vision systems face significant challenges in identifying such dynamic gestures. Due to its limited frame rate, usually 30 to 60 frames per second, motion blur and insufficient sampling are easily generated during high-speed hand movement, resulting in the loss of key action features and recognition delay of more than 50 milliseconds. According to the research of IEEE Transactions on Pattern Analysis and Machine Intelligence in 2023, the frame-based system significantly reduces the accuracy of gesture recognition under complex lighting and fast action conditions.

[0152] Using frame cameras to perform gesture recognition has multiple technical bottlenecks. First, its fixed frame rate mechanism leads to insufficient sampling at high speed, making it difficult to accurately capture the start and end points of key actions. Second, motion blur caused by rapid movement severely reduces image clarity, making it difficult to identify hand contours. Third, frame-based systems continuously collect full-frame images, containing a large amount of redundant background information, resulting in large data volume and low computational efficiency. Fourth, in low light or strong light changing environments, imaging quality drops sharply, affecting recognition stability. Most importantly, cameras capture complete image pictures, which easily reveal personal identity information, not meeting the compliance requirements of privacy-sensitive scenarios.

[0153] In contrast, event sensors have extremely high temporal dynamic response capability, with an equivalent frame rate of up to tens of thousands of frames per second, and a detection error of less than 2 milliseconds for the start and end points of actions, enabling accurate capture of subtle motion changes in hand edges and generation of high-density event clusters. Experimental data shows that the average recognition accuracy of gesture based on event stream deep learning model on public data sets such as N-CALTECH101 and DVS-Gesture is more than 98.5%, significantly better than similar algorithms based on RGB video. Its low latency and high robustness are particularly suitable for safety-critical scenarios.

[0154] Multiple sensors synchronously collect event streams triggered by hand movements and transmit raw event data to edge computing devices or cloud servers through Wi-Fi 6 or 5G networks for centralized processing. To improve recognition reliability, the system uses a multi-view event fusion strategy to align event streams from different locations using a spatiotemporal registration algorithm to construct a more complete three-dimensional hand movement trajectory, effectively alleviating the occlusion problem under single-view.

[0155] At the processing end, the event stream is organized into event frames or voxel grid structures, input into convolutional neural network or graph neural network models for classification and recognition. The models used, such as EventCapsNet or Spike-Driven CNN, have lightweight characteristics and can efficiently run on resource-constrained devices. The recognition process can be completed locally on edge devices, such as industrial computers equipped with Qualcomm AR1 Gen 1 or Sunway RK3588, achieving low-latency response and local data processing. It can also be trained and inferred on cloud servers, supporting cross-scenario knowledge transfer and model optimization.

[0156] The recognition results generate standardized structured instructions, such as SOS requests, emergency alerts, or system shutdown commands, and are pushed to user terminal devices, such as smartphones, smart speakers, or care dispatch platforms, through encrypted channels. The processed feedback information is presented in non-imaging form, such as dynamic human silhouette animations or symbolic icons, allowing users to understand the current state through these abstract visualizations, while the original event stream is not stored or transmitted back.

[0157] The system can be deployed in both private home environments and public areas, supporting edge-independent operation and cloud-edge collaboration modes. The event-driven mechanism only outputs data when it changes dynamically, with communication and computing loads only one to five hundredth of traditional video surveillance, significantly reducing bandwidth usage and energy consumption, making it suitable for long-term continuous operation.

[0158] Figure 9 A deployment scheme for storing in a local area network in an embodiment of the present application is shown, which emphasizes the security of local closed-loop processing and access of data. A visual sensor 201 is installed in the monitoring area to collect human activity information in the environment. The sensor uses an event-driven mechanism and only outputs data when detecting brightness changes, without generating traditional image frames, thereby ensuring privacy from the source.

[0159] The visual sensor 201 is connected to the local storage device 901 through PoE, i.e., Ethernet power supply. This connection method not only transmits data but also powers the sensor, simplifying the wiring process and improving deployment flexibility. All raw event stream data is processed locally, and desensitized human animation and structured behavior metadata are generated after edge computing module analysis. These results are directly stored in the local storage device 901, which can be a NAS device or a smart gateway with storage function.

[0160] User terminals 802, such as smartphones, tablets, or computers, access the same network environment through local area network WiFi, without the need to go through external Internet to directly access the local storage device 901. After authorization, users can browse the de-identified human-like animations in the history record and view the behavior trajectory and action evolution process in a specific time period. Since all data is saved in the user's own device and not uploaded to any remote server, the risk of data leakage in public network transmission is avoided.

[0161] Figure 10 A deployment schematic of the hybrid transmission architecture in the embodiments of the present application is depicted, which aims to combine the security of local storage with the convenience of cloud services, providing a dual-channel data flow solution. First, the visual sensor 201 is responsible for collecting image information in the environment, using an event-driven mechanism to output data only when key activities are detected, in order to protect privacy and reduce unnecessary data processing.

[0162] The visual sensor 201 is connected to the local storage device 901 through PoE, completing preliminary data processing and de-identification operations in the local environment, generating processed human-like animations and behavior metadata, and saving them in the local storage device. Such a design ensures that raw sensitive data does not leave the user's local area network, thus ensuring the privacy and security of the data.

[0163] At the same time, in order to meet the needs of remote access, the local storage device 901 is configured with a network interface such as WiFi or traffic card, which can upload the processed de-identified data to the cloud server 801. Encryption transmission technology is used here to ensure the security of data during network transmission. The cloud server 801 plays the role of a data transfer station, which not only stores the information uploaded from the local device, but also manages and optimizes these data to efficiently respond to the requests of the user terminal 802.

[0164] User terminals 802, such as smartphones, tablets, or personal computers, connect to the cloud server 801 through the Internet. Once authorized, users can view the de-identified human-like animations and other related information uploaded by the local storage device at any time and anywhere. This setup provides users with a flexible and secure way to monitor and manage the information they care about, maintaining contact with home or workplace no matter where they are.

[0165] The entire system constructs a dual-protection data flow: on the one hand, all sensitive data is processed and stored locally, ensuring the highest standards of privacy protection; on the other hand, remote access and management of data are achieved through cloud services, enhancing the system's availability and flexibility. This hybrid transmission architecture is particularly suitable for application scenarios that require both high levels of privacy protection and remote monitoring capabilities, such as smart homes, enterprise security, or public facility monitoring.

[0166] In one comprehensive embodiment of the present invention, the smart community system is deployed in multiple key scenarios within a large community, including home-based elderly residents, community nursing homes, and community hospitals. Based on event-driven visual sensing technology, and combined with edge intelligence and a hierarchical data management architecture, the system enables non-intrusive monitoring and service support of the behavioral status of the elderly population while protecting user privacy.

[0167] In the residential environments of elderly people living at home, each household deploys an independent privacy computing sensor device. This device features a unibody, sealed housing design and integrates Sony's EVS series event-based visual sensor chips (such as the IMX636) and edge computing chips with AI processing capabilities, such as the Rockchip RK3588. The sensor outputs a microsecond-level timestamped event stream only when it detects pixel-level brightness changes, avoiding the generation of facial or body posture information by not acquiring traditional image frames. All raw event data is processed locally on the device, executing lightweight neural network models to perform... Figure 4 The process shown extracts motion features of human body parts and generates desensitized humanoid animations and structured metadata.

[0168] The processed data is uploaded directly to a private cloud server via home Wi-Fi or a built-in 4G data SIM card. Because the original image remains on the device, and the uploaded content consists only of anonymized graphics and lightweight binary packets (such as…), the data is processed automatically. Figure 5 As shown), even if the data is intercepted during transmission, it cannot be reconstructed into a recognizable personal image. This direct-to-cloud architecture (such as...) Figure 8 (As shown) This allows children or caregivers to remotely monitor the activity status of elderly family members through authorized terminals, promptly detect abnormal behaviors such as falls or prolonged periods of inactivity, and enable rapid response. Furthermore, the system does not rely on the complex configuration of home routers, lowering the deployment threshold.

[0169] In community-based elderly care facility scenarios, the system adopts a deployment mode with centralized storage within a local area network (e.g., Figure 9The multiple independent sensors are connected to a central storage device inside the nursing home through PoE (Power over Ethernet) cables. The device can be a dedicated server, a NAS storage system, or an industrial computer equipped with large-capacity hard drives, deployed in the on-site computer room and maintained by the management personnel. This localized storage architecture ensures that sensitive data does not leave the institution, meeting the strict requirements of the nursing home for data sovereignty and privacy protection. At the same time, with the authorization of the regulatory authority, the system can selectively upload some aggregated statistical information (such as the activity frequency of each bed, the number of night-time bed exits, etc.) to the government regulatory platform or cloud service system (such as Figure 10 ). These data can be used to prove bed occupancy, support the nursing home's application for government subsidies or pass the compliance check without providing any original images or individual identity information. This dual-channel design meets the regulatory transparency requirements while achieving data minimization and de-identification.

[0170] In the application of community hospitals, the system further strengthens data isolation and security protection. Independent sensors are connected to the hospital's internal server through PoE or wired network, and all data processing and storage are completed within the hospital's internal network without connecting to any external Internet or uploading to the cloud. The server can be connected to the hospital's HIS (Hospital Information System) or electronic medical record system through a dedicated internal network or a physically isolated dedicated line to realize the structured input of behavior data. For example, the system can automatically record the number of postoperative patient bed exit activities, rehabilitation training duration, and other indicators, and write them into the medical record system in the form of desensitized graphics and metadata to assist medical staff in evaluating rehabilitation progress.

[0171] This embodiment builds a flexible, scalable, and privacy-friendly smart community monitoring system through the three deployment modes shown in Figure 8 , Figure 9 , Figure 10 In the home scenario, the direct-to-cloud architecture ensures the real-time and convenience of remote monitoring; in the nursing home, local storage and selective uploading achieve a balance between privacy protection and regulatory compliance; in the hospital, completely isolated internal network deployment ensures the security of medical data.

[0172] In a specific embodiment, a privacy computing visual monitoring technology scheme for a home security call center is suitable for Figure 8 , Figure 9 and Figure 10The three system architectures shown are designed to address the privacy leakage risk caused by image exposure in traditional video monitoring in remote care scenarios, while overcoming the technical shortcomings of non-imaging sensors in intuitively expressing human behavior states. The system takes the visual sensing unit equipped with Sony Prophesee event-type image sensors as the core, and processes the raw event stream locally or remotely into desensitized visual data for call center agents to observe and judge the activity state of the elderly in real time, such as whether they have fallen, whether they have been stationary for a long time, whether they are undergoing rehabilitation training, and so on, so as to initiate emergency response in a timely manner.

[0173] In this embodiment, the processed raw event stream can be presented in three different levels of visualization modes to adapt to different lighting conditions and observation needs. The first is the brightness mode, in which the system maps the event density to the gray intensity, and the more intense the motion, the higher the brightness of the area, and the stationary area remains dark. In this mode, the human body appears as a dynamic light spot profile, and the agent can quickly judge the behavior state by the overall brightness distribution and motion trajectory, such as a sudden spread of high-light area when falling, and a continuous movement of brightness center when slowly getting up. This mode has small computational complexity and is suitable for most daily scenarios, and it preserves privacy without showing any structural details.

[0174] When the ambient light is complex or the human motion is subtle, making it difficult to identify in the brightness mode, the system can switch to the second color mode. This mode identifies the active areas of each part of the human body through a clustering algorithm and assigns different color identifiers, such as blue for the head, green for the upper limbs, red for the lower limbs, and yellow for the torso. The color coding is based on motion consistency analysis and spatial connectivity judgment, and does not rely on any image reconstruction or AI model. The agent can intuitively identify the position and motion direction of the limbs by the color distribution, such as determining whether the old person is trying to grasp the handrail or standing unsteadily on one leg, to more accurately assess the fall risk or the normativity of rehabilitation movements. This mode significantly improves the accuracy of behavior interpretation while maintaining the non-imaging characteristics.

[0175] If the first two modes still cannot clearly present the key details, the system can further switch to the third gray-scale detail mode. This mode adds local contrast enhancement and edge preservation filtering to the brightness, generating a high-resolution dynamic contour map by accumulating event streams from multiple time windows. This contour map only shows the outer edge of the human body and the main limb boundaries, without containing identifiable information such as faces and clothing textures, and all pixel values are calculated, making it impossible to reverse to the original image. This mode is suitable for low-light or heavily obstructed scenarios, helping agents confirm subtle movements such as finger tremors, slight convulsions, and so on, providing support for medical-level judgments.

[0176] The conversion of the above three visualization modes is triggered by the call center staff according to the actual observation needs, and all modes do not display real images, ensuring privacy protection throughout. More importantly, the location of the visualization processing can be flexibly deployed according to the system architecture. In Figure 8 In the direct cloud architecture shown, the sensing unit completes the conversion of the event to the visualization data locally, and only uploads the desensitized graphic data stream to the cloud server, and the agent views it in real time through the Web client. This way ensures that the original event stream does not leave the device, and even if the cloud is attacked, sensitive information cannot be obtained.

[0177] In Figure 9 In the local closed-loop architecture shown, multiple sensing units transmit the original event stream or preliminary processing results to the local server, and the server uniformly performs visualization rendering. This mode is suitable for community care institutions, and administrators can directly access the server within the local area network to view the real-time activity status of each room. Since the data is stored and processed locally, the risk of public network transmission is avoided, meeting the strict requirements of the institution on data sovereignty.

[0178] Exemplarily, the event-type image sensor works in an asynchronous manner, and only outputs event data when the pixel senses a change in illumination. The data stream output by it is highly sparse and uneven in time. When there is rapid movement of personnel, drastic changes in illumination, or multiple targets active at the same time in the monitored scene, the sensor will generate a large number of dense event data in a very short time, forming a transient data peak. Since this type of sensor does not use the traditional frame output mechanism, its data generation rate is not limited by the fixed frame rate, resulting in a data volume that may far exceed the transmission bandwidth of the communication interface and the processing capacity of the receiving end in a unit of time. This transient overload is easy to cause the overflow of the receiving buffer, and then cause event data loss, timestamp disorder or transmission delay. The direct consequence is that the original motion information is incomplete, affecting the accurate reconstruction of the motion trajectory, which may lead to behavior recognition misjudgment, missing of action fragments, or visualization display stuttering, jumping. If the system cannot recover in time, it may also cause processing thread blocking or resource exhaustion, reducing the overall stability and reliability of the system, and ultimately affecting the continuity and usability of the privacy computing visualization function.

[0179] In one possible implementation, the application aims at the event camera's event instant overload problem caused by the mismatch between data output characteristics and transmission mechanism in the visualized privacy computing system, and proposes a scene-adaptive MIPI transmission parameter dynamic regulation method. The system uses the IMX646-AAMR-C produced by Sony as an event-driven visual sensor, and the AK3918 AV100 produced by ANYKA as an edge computing processor to build a complete localized perception architecture with low delay and high privacy. The IMX646-AAMR-C transmits asynchronous event streams to the AK3918 AV100 through the MIPI CSI-2 interface, and the latter is responsible for receiving, caching, reorganizing and AI analyzing the event data, and finally generating optical flow trajectories that can be used for behavior recognition. The entire process does not output any traditional image frames, only abstract event data containing spatial coordinates, polarity and timestamps are transmitted, thereby realizing the visual modeling of user activities without revealing visual privacy.

[0180] However, in actual tests, it is found that there is an essential difference between IMX646-AAMR-C and traditional RGB image sensors in the MIPI transmission mechanism, which is the root cause of the overflow of the received data. The traditional RGB sensor outputs a fixed number of pixel data in each frame period, and the time interval from FS (Frame Start) to FE (Frame End) and the data volume are both constant values, so the system can pre-allocate buffer resources and plan the processing period accordingly. Although the time period from FS to FE of IMX646-AAMR-C is set to a fixed value by the register mipi_frame_period, for example, it is configured as 0x000007D0 corresponding to about 2 milliseconds, but the number of events generated in this period depends entirely on the scene dynamics, showing a highly sparse and bursty feature. In a static scene, only tens of events may be generated in a FS-FE period, while in a scene with multiple people moving quickly or with dramatic changes in lighting, the number of events can increase to tens of thousands in milliseconds. This asynchronous output mode of "fixed time and dynamic data volume" makes it difficult for the receiver to adopt the traditional frame processing strategy, and if not optimized, it is easy to cause MIPI receiving buffer overflow, event loss or processing delay in high activity scenes, thereby affecting the integrity of subsequent optical flow reconstruction and the accuracy of behavior recognition.

[0181] To solve the above technical problems, a complete asynchronous event flow receiving and scheduling mechanism is designed for AK3918 AV100 terminal. First, the hardware level ensures that IMX646-AAMR-C and AK3918 AV100 MIPI CSI-2 interface are electrically compatible: both support 1.5 Gbps / Lane transmission rate, adopt 2 Lane configuration, differential impedance matching is 100Ω, and the power voltage is 1.8V LVDS level, without additional level conversion circuit. When PCB wiring, the MIPI differential pair length is matched within ±100μm, the terminal resistance 100Ω is connected in parallel to the receiving end, and the wiring length is not more than 15cm, to ensure signal integrity. Second, the IMX646-AAMR-C is configured through I2C bus, and the key parameters include: the register 0x0000B020 (mipi_packet_size) is set to 0x00001000 (4KB), to avoid the default 16KB packet being too large to cause delay; the register 0x0000B021 (mipi_packet_timeout_enable) is set to 1, and the timeout value is set to 50μs, to ensure that short packets can be sent in time when the activity is low; the register 0x0000B02c (mipi_frame_period) is set to 0x000007D0 (2ms), to ensure stable output period.

[0182] On the AK3918 AV100 side, the MIPI receiving module is configured as 2 Lane, 1.5 Gbps, external clock mode, and the DCKP / DCKN synchronous clock is provided by the IMX646. In the data type mapping, 0x00 is defined as frame start (FS), 0x01 is defined as frame end (FE), and 0x02 is defined as event data (EVT), to ensure correct protocol parsing. The receiving end enables the DMA direct memory access function to write the event data stream into the ring buffer in the external SDRAM, and the buffer capacity is not less than 64 KB to prevent burst data overflow. At the same time, the processor is built-in with a high-precision timer with 1 μs resolution, which is used to calibrate the local clock and align with the event timestamp, to ensure the timing continuity of cross-packet events. In the firmware layer, the mixed mode of interrupt and polling is used to handle the event arrival, to avoid frequent interrupt overhead, and the event reordering algorithm is implemented to ensure the accuracy of optical flow calculation. According to the system time and real-time event density, the MIPI transmission parameters are automatically adjusted. When the system time is in the night period from 22:00 to 06:00 of the next day, or the total number of events detected within 1 second is less than 500, it is determined as a low activity scene, and the edge processor adjusts the mipi_packet_size to 1024 bytes and the mipi_packet_timeout to 30 μs through the I²C interface, to ensure fast response. When the system time is in the daytime regular period from 06:00 to 22:00, or the event rate is maintained between 500 and 5000 per second, the system enters the medium activity mode, and the parameters return to 4 KB packet size and 50 μs timeout. When the system enters the preset high activity period such as the family gathering period from 18:00 to 20:00, or the event rate is continuously detected to be more than 5000 per second for more than 100 milliseconds, the system switches to the high activity mode, increases the mipi_packet_size to 8192 bytes, and extends the mipi_packet_timeout to 100 μs, to improve the throughput efficiency. The switching process is completed by the edge processor monitoring the event flow density in real time and combining the system real-time clock, and the processor dynamically writes the registers of the event camera according to the current state, to realize smooth switching of parameters and ensure that the event flow is not interrupted and lost during the switching process.

[0183] The system supports a user to preset a scene label such as a "night monitoring" or a "party mode" through an application interface, and binds the label with a time interval, so that the configuration flexibility is improved. During parameter adjustment, if I2C communication fails, the system automatically reverts to the last stable configuration, so that the operation reliability is ensured. All received event data are written into a ring buffer, and are reorganized into a visual light flow sequence in a timestamp order, and are used for subsequent behavior identification and trajectory analysis. The method not only solves the data congestion problem caused by event bursts, but also ensures stable output of light flow data through a dynamic adaptation mechanism, so that the system can continuously generate high-quality motion trajectories under different illuminations and activity intensities, and effectively supports visual application scenarios such as fall detection, personnel loitering and abnormal behavior identification.

[0184] The technical solution fully utilizes the NPU and ISP capabilities of AK3918 AV100 without relying on high-performance GPUs or special AI acceleration chips, and reduces the overall computing power requirement and power consumption level in combination with the sparsity advantage of event data itself. The IMX646-AAMR-C and AK3918 AV100 used are mature industrial-grade chips in mass production, and the supply chain is stable and the cost is controllable. The hardware material cost of the whole solution is close to the mainstream 1080p high-definition camera level, and no additional large-capacity storage, high-speed interface or heat dissipation structure is needed to realize the intelligent visual perception function with privacy protection capability. Therefore, while ensuring the smoothness of the visual processing, the hardware cost deployment is low, so that the visual solution based on privacy computing can be widely applied to cost-sensitive consumer scenarios such as home monitoring, elderly care, child safety and the like.

[0185] In Figure 10 In the hybrid architecture shown, the system supports a hierarchical authorization mechanism. In the daily state, the sensing unit only outputs the lowest detail brightness mode data for routine monitoring by the on-duty personnel. When an abnormal event occurs or external inspection is accepted, the color or grayscale detail mode can be temporarily enabled after authorization, and the visual data of the related period is encrypted and uploaded to the supervision platform. The uploaded content is only behavior animation, and does not include original events or images, which not only meets the compliance review requirements, but also prevents the abuse of long-term high-detail data.

[0186] The embodiment solves the dilemma between privacy and readability of traditional monitoring through three switchable visualization modes. The brightness mode guarantees basic observation, the color mode improves judgment accuracy, and the grayscale detail mode deals with complex scenes, all of which are directly generated based on event streams without relying on image reconstruction. At the same time, flexible deployment of the processing location enables the system to adapt to different environments such as families, communities and institutions, and meets the multi-level needs from personal privacy to institutional compliance. Call center personnel do not need to face real images to accurately judge the state of the elderly, which not only reduces the psychological burden, but also improves the response efficiency.

[0187] In the home security call center embodiment, especially for sensitive spaces in the patient's living area where privacy protection requirements are extremely high, such as bathroom, toilet, bathroom, etc. Scene, all-weather, non-invasive, non-imaging safety monitoring is realized. These areas cannot deploy conventional cameras due to the involvement of body exposure and private behavior, resulting in a blind area in safety monitoring. However, the invention relies on the physical characteristics of the Sony Prophesee event-type image sensor, which only responds to the dynamic changes in pixel-level light intensity, and outputs a 64-bit sparse event stream containing spatial coordinates (x, y), timestamp (t) and brightness change polarity (polarity). Since the sensor does not perform full-frame scanning and does not collect static background information, it cannot form a complete image and does not have the ability to record identifiable biological information such as facial features, skin color, and body outline. In the bathroom or bathroom environment, even if the monitored person is in a naked state, the sensor only captures a sequence of local light intensity change events caused by limb movement. These data are essentially non-imaging spatiotemporal pulse signals and do not contain any visual images.

[0188] On this basis, the system performs dynamic event focusing (DEF) and event stream mapping (ESM) processing through the edge computing module, converting the original event stream into three levels of desensitization visualization output. In all modes, the system only presents the overall posture and action trend of the moving human body, which is manifested as follows: in the brightness mode, the human body is displayed as a dynamic light spot composed of high-light pixels, and its shape continuously evolves with the action, such as a vertical ellipse when standing and quickly spreading into a horizontal distribution when falling. In this mode, there are no structural details, and only the motion intensity and center of gravity displacement are reflected; in the color mode, the system divides the human body into four logical regions of head, upper limbs, lower limbs and torso through spatial clustering and motion consistency analysis, and assigns fixed color coding (such as blue, green, red and yellow) to each region. These color blocks only represent the approximate orientation and relative motion relationship, do not reflect the limb length proportion or joint bending angle, and all regions are regular geometric shapes (such as circles and rectangles), avoiding the formation of real human body outlines; in the gray-scale detail mode, the system can generate a simplified contour line graph based on event accumulation to display the outer edge of the human body and the main limb boundaries. The contour is blurred and simplified, with a resolution limited to 32x32 pixels, which cannot distinguish fingers, faces or private parts, and is only used to determine whether it is stationary for a long time or has an abnormal posture (such as curling or falling). All the above visualization results are packaged through a lightweight binary protocol and do not contain raw event data, and cannot be reversed to any form of image.

[0189] More importantly, the system is designed to track and display only moving targets by default, and areas that are static for more than 3 seconds are automatically removed from the screen, ensuring that when a person is in a relatively static state during the process of using the toilet or taking a bath, their presence is greatly weakened, further reducing the psychological discomfort. This mechanism not only greatly reduces the amount of data and transmission bandwidth, keeping the background monitoring screen simple and clear, avoiding irrelevant interference, but also significantly improves the input quality and efficiency of AI analysis. The deployment of the system significantly reduces the dependence on manpower for managers in daily patrol. In the traditional care mode, nursing staff need to enter the patient's room regularly for on-site observation to confirm their safety status. This process not only consumes a lot of human resources, but also may affect the patient's independence and psychological comfort due to frequent disturbances. Especially in privacy-sensitive areas such as bathrooms and toilets, manual patrol is difficult to continue and can easily cause resistance from the cared-for. The invention enables remote, non-intrusive, and continuous activity state perception, allowing managers to grasp real-time dynamics without frequent room entry, thereby significantly reducing patrol frequency and single observation time, effectively reducing labor costs and operational burden.

[0190] In the actual work of the nursing team, the desensitization visual output of multiple sensing units can be unified to the central monitoring platform of the nurse station to form a comprehensive view of multiple screens displaying in parallel. Each screen only presents the desensitized motion profile or color-coded body posture, without image information that can identify the identity, ensuring that privacy protection principles are strictly followed while monitoring centrally. Nursing staff can observe the activity status of multiple rooms simultaneously on one screen, quickly identify abnormal behaviors such as falls, long-term static or violent shaking, and thus achieve "at a glance" efficient management. This centralized monitoring mode not only improves response speed, but also allows patients to freely move in private spaces, reducing psychological stress caused by human intervention and enhancing their sense of dignity.

[0191] The system design fully considers the expansion needs of intelligent management. All desensitized visual data streams can be output to local edge computing devices or cloud AI analysis platforms through standardized interfaces (such as JSON or Protobuf format). Since the data itself is a structured or semi-structured motion feature sequence (such as target position, motion vector, and part state), compared to the original video stream, it has a small data volume, clear semantics, reduces the input complexity of the model, and improves analysis efficiency and accuracy. For example, edge computing models can judge behavior patterns in real time based on these data, predict fall risks, and generate warning signals; cloud platforms can perform clustering analysis on long-term behavior data to assist in developing personalized care plans or evaluating rehabilitation progress.

[0192] The nursing team can not only respond to emergencies in real time, but also make proactive interventions based on the behavior trend reports provided by the system, further optimizing resource allocation. For example, the system can automatically identify a patient's regular night-time frequent night-time, prompt nursing staff to adjust the lighting or arrange for an escort, and thus improve the quality of care without increasing the number of staff. This is a "low-invasive, high-intelligent" management mode.

[0193] The present application provides another possible implementation, which is suitable for home nursing service scene, aims to solve the problem of serious privacy leakage caused by using high-definition camera to record service behavior in current nursing process. In existing practice, nursing staff generally wear high-definition cameras or smart glasses for whole-process video recording to ensure that the service process is traceable, verifiable, and serves as a service completion certificate and liability exemption basis, especially when dealing with operations involving private parts of the body such as bed sore debridement and wound dressing, such video data becomes the key evidence for service quality supervision and long-term care insurance reimbursement audit. However, the traditional high-definition video recording method inevitably collects and stores a large amount of sensitive pictures containing faces, body features and private parts. Once these data are leaked or misused, it will cause irreversible infringement on the dignity and privacy of the service object, and also increase the data security compliance risk of the institution.

[0194] In this embodiment, a typical application scenario is to integrate a privacy computing unit based on a miniature event-based vision sensor (DVS chip) into a wearable device, especially for smart glasses, head-mounted recording devices, and chest-mounted recording devices. The smart glasses are a frame structure worn above the user's ears, with the temples spanning the sides of the nose bridge. The appearance is similar to conventional optical glasses or sunglasses, and the front frame or nose bridge connection is integrated with a miniature optical lens and a DVS chip for continuously sensing the dynamic changes of the scene in front of the user at the first visual angle. The device realizes long-time wearing through lightweight design and is suitable for daily life recording, auxiliary visual navigation, and other scenarios. The head-mounted recording device is a separate device fixed to the user's head, usually in the form of a headband, headband, or helmet accessory. The sensing unit is installed in the forehead area, and the lens is directed forward to ensure stable acquisition of event stream data within the field of view during movement, operation, or inspection. It is suitable for security inspection, industrial operation recording, and other application scenarios that require the liberation of both hands. The chest-mounted recording device is a small module that can be fixed to clothing or a shoulder strap. It is usually installed on the user's chest or shoulder through clips, magnetic attraction, or straps, and the lens is directed forward to capture the user's overall behavior in a third-person perspective. It is suitable for law enforcement recording, nursing staff work trace recording, and child activity monitoring, and other scenarios that focus on behavior process recording. The three types of devices all use built-in embedded processors to achieve local privacy computing and output desensitized visual data streams, ensuring privacy protection at the source. By replacing traditional image acquisition modules with low-power, small-size event-based image sensors (such as Prophesee Gen4 or equivalent chips), the technical path of eliminating complete image frame generation from the perception source is realized. The DVS chip can be embedded in the smart glasses frame, worn on the chest of nursing staff, or fixed to the designated position of work clothes. It has a visual angle coverage capability comparable to conventional high-definition cameras (horizontal field of view ≥ 70°). Its output is only an asynchronous sparse event stream (x, y, t, polarity), which only responds to pixel-level light changes, does not perform full-frame scanning, and does not capture static background information. Therefore, it cannot reconstruct visual images that can identify individual identities, fundamentally avoiding the risk of collecting sensitive images.

[0195] Further, multiple such privacy-computing wearable devices can be networked through a local area network or wireless communication protocol (such as Wi-Fi 6, Bluetooth Mesh, or 5G NR) to form a distributed sensor network. Each device acts as an independent data collection node, synchronously uploading its generated de-identified visual data stream to a local edge server or central management platform, enabling spatio-temporal alignment and fusion analysis of multi-perspective behavioral data. For example, in a multi-person collaborative care scenario, multiple nurses wearing smart glasses can simultaneously access the system, and their visual line movements and operation trajectories are uniformly recorded and mapped into non-imaging dynamic identification sequences, supporting panoramic restoration and quality traceability of the service process. This networked deployment mode not only expands the spatial coverage, but also enhances the integrity verification capability of the service process, suitable for institutional management, team operation supervision, and cross-regional collaborative service in complex scenarios, while maintaining the entire link without raw image transmission, ensuring that user privacy is always under control.

[0196] In terms of technical implementation, the system adopts a high-performance heterogeneous computing architecture commonly used in mainstream AI glasses, integrating Qualcomm AR1 Gen 1 as the main control SoC, supplemented by low-power co-processors such as BES2700 from Handpoint Technology. This solution draws on the hardware design paradigms of products such as Meta Ray-Ban smart glasses, Xiaomi AI glasses, and Rokid Glasses, ensuring the feasibility of the system in terms of computing power, power consumption, and integration. Among them, Qualcomm AR1 Gen 1 uses a 6nm process, integrating multi-core CPU, GPU, third-generation Hexagon NPU, and dedicated ISP modules, with powerful edge AI processing capabilities; Handpoint BES2700 is responsible for low-power audio and sensor fusion tasks, optimizing overall energy efficiency. Under this architecture, the system runs Dynamic Event Focus (DEF) and Event Stream Mapping (ESM) algorithms, which convert raw event streams into de-identified visual data in real time.

[0197] The system further integrates an eye tracking module, which detects the wearer's pupil position and visual line direction in real time by arranging a near-infrared miniature camera and an infrared LED array on the inside of the glasses. When a caregiver performs high-precision operations, such as debridement, medication, or bandaging of a pressure ulcer, their visual line will naturally focus on the operation area. The system dynamically adjusts the event processing priority of the DVS sensor based on eye tracking data, only performing high-density event aggregation and detail enhancement in the foveal region, while maintaining basic motion contour display in the peripheral field of view. This "eye movement guided adaptive visualization" mechanism achieves intelligent allocation of computing resources and data output, significantly reducing the overall data volume.

[0198] In practical applications, the system only generates higher-detail desensitized contours or color-coded blocks in the area where the caregiver actively focuses on, while the rest of the area remains in minimalist light spots or complete silence. For example, when the caregiver bends down to observe the patient's leg bedsores, the system automatically enhances the event resolution in that local area, generating visual features that can be used to determine whether the medication is evenly applied and the action is standardized; when the line of sight moves away, the area immediately returns to low-detail mode. This on-demand enhancement mechanism greatly reduces the generation and transmission of invalid data, allowing limited local storage space to support longer service records, reducing storage costs and cloud bandwidth pressure.

[0199] The desensitized visual data contains three configurable modes: in the brightness mode, human motion light spots are presented to determine the relative position between the caregiver and the patient and the interaction state; in the color mode, the head, torso, and limbs are identified by fixed color blocks to facilitate the identification of operation action direction (such as bending, reaching, and wiping); in the grayscale detail mode, simplified contours are provided to confirm the execution process of specific care actions (such as hand approaching the wound area). All outputs are non-imaging graphics that cannot be restored to the original picture, and the system defaults to only track moving targets, with static areas automatically disappearing, further reducing data sensitivity.

[0200] The desensitized visual data can be uploaded in real-time to the local gateway or cloud server through Bluetooth 5.3 or Wi-Fi 6E, or temporarily stored in the ePOP encapsulated storage module (integrating LPDDR5 memory and UFS flash memory) built-in the device, and then uploaded uniformly after the service is completed. The uploaded content does not contain any original event stream or image frame, only contains structured behavior feature sequences (such as target ID, position coordinates, motion vector, and posture encoding), supports automatic association with service work order systems, and serves as a digital certificate of the service process. Insurance audit agencies or regulatory platforms can view these desensitized animations through dedicated clients to verify whether the service is completed on time and the operation is standardized, for example, to confirm whether there are continuous hand movements and patient position coordination in bed sore care, to determine the authenticity of the service, effectively preventing fraudulent behaviors such as false clock-in, proxy service, and the like.

[0201] The beneficial effects of the present embodiment are significant. First, the privacy protection capability reaches the highest level: cutting off image generation from the perception source, ensuring that patient body features, facial information, and private parts are not recorded, and improving user trust in home service. Second, meet the compliance and audit requirements: desensitized visual data can be used as legal and effective service evidence to support long-term care insurance reimbursement audit and service quality traceability, which not only protects the rights and interests of service providers, but also maintains the supervision rights of the payment party. Third, reduce data security risks and storage costs: as the data volume is only 1% or less of traditional video, and the non-key area data is further compressed through eye movement focusing mechanism, the operation and maintenance overhead and data leakage risk are significantly reduced. Finally, improve the value of clinical support: the system can output standardized behavior logs for doctors to analyze patient body position changes, pain reactions (such as sudden movements), and other information to assist in developing personalized treatment plans.

[0202] In another possible embodiment, the present application is applied to an intelligent monitoring system of a community comprehensive elderly care service center, covering functional areas such as public activity areas, corridors, entrances and exits, public toilets, and shower room outer rooms. The project requirements require strict protection of the personal privacy of the elderly in scenes involving body exposure or private behavior while ensuring public safety and emergency response capabilities, avoiding visual image information collection and dissemination.

[0203] The system uses a hybrid visual perception architecture, flexibly configures differentiated data collection methods according to the functional attributes and privacy sensitivity levels of different areas. In public areas such as halls, corridors, and entrances, standard high-definition network cameras are deployed to support 1080P and above resolution video collection for regular security monitoring, personnel access records, and incident backtracking. These cameras are connected to the original video management platform, and the data is independently stored in a dedicated video storage server after encryption, and the access permission is controlled by the role, and only authorized security personnel can access the data after two-factor authentication.

[0204] In high-privacy-sensitive areas, including public toilet stall exterior monitoring areas and shower room outer rooms, the system no longer uses traditional imaging cameras, but connects to privacy computing visual sensing units based on event-type visual principles. The unit only responds to changes in scene brightness and outputs asynchronous event streams, and completes human motion detection, pose estimation, and behavior semantic recognition in the local edge computing module to generate non-imaging behavior representation data such as "enter", "stand", "sit down", "fall down", and "long time no movement" structured labels. The entire processing process does not form any frame image and does not rely on image reconstruction technology, and the output content is pure data stream.

[0205] The data of the two types of devices is path-separated after collection. The video data of the high-definition camera is transmitted through the RTSP protocol, encapsulating complete pixel information, and stored in a separate video database. The behavior data of the privacy computing unit is uploaded to the behavior analysis server in a lightweight JSON format through the HTTP or MQTT protocol, with fields including timestamp, anonymous number, behavior category, confidence, spatial position coordinates, and duration, and does not contain any information that can be restored to a visual image. The two sets of data are stored in different physical storage media or logical partitions, and do not cross each other, ensuring the purity of the privacy data chain.

[0206] Although the data is stored separately at the bottom layer, the system designs a unified operation interface at the upper layer application to meet the real-time grasp of the overall environment by management personnel. The central monitoring platform supports a multi-view fusion display mode, which can display the video screen of the public area and the behavior state information of the privacy area on the same screen. The specific presentation methods include a grid-like layout, such as displaying real-time video thumbnails of the public area on the left side and displaying behavior state icons from the privacy unit in the corresponding area on the right side, such as a green dot representing "normal activity", a red triangle representing "falling alarm", and a blue horizontal line representing "sitting posture stay". It can also be configured to switch the screen at a custom time, and the system automatically rotates the state information of different areas at a preset time interval.

[0207] All information from the privacy area is presented in a symbolic, semantic, and non-pixelated manner. The system does not perform image decoding, does not generate intermediate frames, and does not cache any temporary pictures for visualization during the rendering process. The behavior state is expressed through vector graphics, color coding, and textual prompts, and its essence is a data-driven dynamic identifier, rather than a degraded or blurred version of a visual image.

[0208] In order to further ensure the data compliance of the privacy area, a dedicated hardware device is connected in series to the network outlet of each privacy computing visual sensing unit. The device has dual network interfaces, connecting between the sensing unit and the local area network, and real-time monitoring of all data packets transmitted outward. Its built-in structure analysis module can analyze the data content and detect whether there are JPEG file headers, MJPEG video stream characteristics, RTSP protocol identifiers, or fixed-period large-size data sequences. Once suspected image frame characteristics are found, the transmission channel is immediately blocked, and a local audible and visual alarm is triggered, with event logs recorded for subsequent auditing.

[0209] This architecture is not limited to a single type of deployment scenario, but can adapt to various spatial combinations, supporting flexible networking from low density to high density. The privacy computing unit and the protection device are independent modules that can be integrated into existing monitoring systems, suitable for various public environments that are highly sensitive to privacy, such as the elderly, medical care, education, etc.

[0210] In the actual deployment of the nursing home toilet, a monitoring scheme is implemented, which replaces the front-end traditional high-definition camera with a new sensing terminal based on event visual sensor without changing any existing monitoring system architecture. The size, installation method, power supply interface and communication protocol of the new sensing terminal are completely consistent with the standard IPC camera, which can seamlessly access the original PoE switch, NVR and display screen system.

[0211] In one exemplary embodiment, the original event data is captured by a Sony IMX646 event visual sensor, directly input to a Rockchip RV1106 master chip through a MIPI CSI-2 high-speed interface, and the complete conversion process from the original event stream to the visualized optical flow image is completed in the chip. Finally, it is encoded into a standard H.264 video stream and output through the RTSP protocol, transmitted to the NVR through the original network and displayed on the monitoring screen. It appears to the outside as a standard network camera that outputs special dynamic pictures, without modification to the back-end platform.

[0212] The physical structure of the device adopts a standard hemispherical shell, supports M4 screw hole ceiling installation, the lens focal length is matched with a 3.6mm fixed focus lens according to the field of view angle of the event sensor, the network interface is a standard RJ45 and supports IEEE802.3af standard PoE power supply, only a single bicolor status indicator is retained for running status prompt, there is no SD card slot, no audio interface, no alarm IO and other redundant design, easy to deploy.

[0213] In the implementation of the core optical flow visualization algorithm, the system adopts a lightweight real-time processing architecture optimized for the embedded computing power of RV1106, does not rely on deep learning frameworks or complex image libraries, and controls all processing procedures within 500 lines of C++ code, with memory occupancy below 50MB and CPU occupancy rate below 30%, ensuring long-term stable operation. The algorithm process includes five key stages:

[0214] The first stage is event cache management. The system continuously receives asynchronous event streams from IMX636, each event containing pixel coordinates x (0~1279), y (0~719), microsecond timestamp and polarity (this scheme ignores polarity). All events are written to a ring buffer, with a buffer capacity of 50 milliseconds, which can accommodate about 5000 event points. Expired events are automatically removed when the frame is generated, ensuring that only the latest valid motion information is processed.

[0215] The second stage is spatial grid mapping. The system divides the 1280x720 resolution picture into 32 columns x 18 rows of rectangular grids, each grid corresponding to a 40x40 pixel area. For each valid event in the buffer, the system calculates the grid index according to its (x, y) coordinates, and performs an atomic increment operation on the event counter of the grid, thereby clustering the sparse event points into a regional density distribution map.

[0216] The third stage is brightness mapping and normalization. The system traverses all 576 grids, counts the maximum event density value in the current frame, and linearly scales all grids based on this value to map the original count value to a gray brightness value of 0 to 255, ensuring maximum dynamic range of the picture and avoiding local overexposure or underexposure, while enhancing the human eye's sensitivity to motion hotspots.

[0217] The fourth stage is time decay animation. To simulate the fade-out effect of motion trajectories in the real world, the system multiplies all grid brightness values by a fixed decay coefficient of 0.85 before generating each frame. In the visualization presentation of the present application, the system dynamically renders historical motion trajectories through a time decay mechanism. Specifically, the recognition results of each body part of the human body are presented frame by frame on the display interface as discrete graphical elements, each containing position, shape, color, and transparency attributes. The system assigns each historical trajectory point a lifespan that is synchronized with the system timestamp. The closer the trajectory point's survival time is to the current time, the higher its transparency and the brighter its color. As its existence time extends, the system gradually reduces its transparency according to a pre-set decay function, and optionally reduces its display size or desaturates its color, until the trajectory point is completely invisible. Finally, a motion path composed of multiple continuous or nearly continuous graphical elements is formed on the display terminal, with the front end corresponding to the latest detection position displayed as high brightness and high contrast, and the rear end corresponding to the position at an earlier time presented as low transparency, weak contrast, or fading effect. This path is not drawn statically or permanently, but is dynamically updated every frame, with old trajectory points disappearing one by one according to a time rule. The visual effect achieved is that the travel route of the moving object appears as a band-shaped trajectory with a clear front end and a fading tail, similar to the continuous light trail left by a moving light source during exposure. This display method preserves the temporal continuity and directional features of the motion, allowing observers to identify behavior patterns such as walking, turning, and staying; at the same time, since historical trajectories automatically decay over time, it avoids forming permanent graphical residues in static scenes, preventing the background area from being covered by accumulated trajectories, thereby maintaining a clear picture and improving the accuracy and readability of behavior interpretation.

[0218] The fifth stage is frame synthesis and rendering. The system creates a 1280x720 gray frame buffer, traverses each mesh, and fills the brightness value uniformly to the corresponding 40x40 pixel area. Finally, a complete gray image frame is generated, which does not contain any real object contour or texture, and only presents the spatial density and time evolution process of motion events. When a person walks, it appears as a continuous moving light point trajectory. When a person falls suddenly, it appears as a light point that spreads instantaneously and gradually stops. Abnormal behavior can be visually identified by monitoring personnel.

[0219] In the encoding output link, the system uses the Rockchip RV1106 built-in hardware H.264 encoding engine to input the above generated gray frame as a YUV420 format brightness component (Y channel), and fills the U / V channel with 128 neutral values to construct a standard YUV input frame. The encoding parameters are configured as Baseline Profile, CBR constant bit rate mode, 25fps frame rate, 1280x720 resolution, and 2Mbps rate, ensuring compatibility with mainstream NVR and controllable network bandwidth. The encoded video stream is sent to the media framework through the Linux V4L2 interface, and then packaged into a standard RTSP stream by a lightweight RTSP service (such as live555 or GStreamer rtspsrc plugin). The URL format is rtsp: / / [device IP]:554 / stream, supports ONVIF protocol discovery, and can be directly added and previewed by mainstream NVR platforms such as Hikvision, Dahua, and Uniview.

[0220] The device only outputs RTSP video stream by default, does not expose raw event data, does not push structured JSON, and does not rely on cloud platforms or new software systems. Users can access the device IP address to enter the simple Web configuration page and adjust network parameters, RTSP port, light flow density threshold, decay coefficient, and other basic settings. All configurations take effect immediately without the need for a restart.

[0221] The deployment process is extremely simple. After power off, the original IPC camera is removed and replaced with this device. After power on, the corresponding RTSP address is manually added in the NVR to immediately see the dynamic light flow picture. No changes are needed to the switch, recorder, display, or rewiring throughout the process, and the process can be completed independently by a caregiver or engineer.

Claims

1. A privacy-preserving computational visual sensing unit, characterized in that, It consists of the following integrated functional modules: The optical signal acquisition module includes an optical lens and an event-type image sensor, which is configured to respond asynchronously to pixel-level illumination changes and output a raw event stream containing spatial coordinates, time information and brightness change polarity, rather than periodic image frames. A data processing interface module is used to output the raw event stream to an external processing device, or to perform event stream analysis locally; A privacy-preserving computation engine, configured on a local or external processing device, receives the raw event stream and generates a privacy-preserving computation visualization data stream. This visualization data stream provides a readable representation of human motion states in the following ways: Transform event streams into sparse point clouds, cumulative contours, or structured morphological representations; Visual identification of different body parts of the human body is achieved based on event density, consistency of movement direction or spatial clustering analysis, without relying on image reconstruction or artificial intelligence recognition models. A data output module is used to transmit the privacy computing visualization data stream to a display terminal or storage device; In this system, no node generates, stores, or transmits any form of complete image frame, and the output visualization data cannot be restored to image information that can identify an individual, thus achieving privacy protection at the source of data collection. The privacy computing visual sensing unit is also configured to: combine the system's real-time clock information and determine the current scene mode based on the activity intensity of the original event stream, and dynamically adjust the communication packet parameters during data transmission according to different scene modes, so as to achieve adaptive encapsulation and transmission management of the event stream.

2. The privacy-focused computational vision sensing unit according to claim 1, characterized in that, The "visual identifier" is achieved by assigning distinguishable visual features to different body parts, including: Color attributes, including hue, saturation, brightness, or transparency; Geometric shapes, including points, lines, surfaces, symbols, or the degree of closure of contours; Dynamic display modes include continuous display, periodic flashing, pulsed on / off, or gradual brightness change; The visual features are assigned based on the analysis results of event density, consistency of motion direction, or spatial relative position.

3. The privacy-focused computational vision sensing unit according to claim 1 or 2, characterized in that, The body parts include at least one of the head, upper limbs, lower limbs, and torso. The edge computing module establishes a correspondence between active event areas and body parts through spatial clustering and motion feature analysis.

4. The privacy-preserving computational visual sensing unit according to any one of claims 1 to 3, characterized in that, In the privacy-computing visualization data stream, at least two body parts are distinguished and displayed using different visual features.

5. The privacy-preserving computational visual sensing unit according to any one of claims 1 to 3, characterized in that, All body parts are displayed using the same visual features, but the visualization data stream still retains semantic information about the parts for subsequent processing or user-defined rendering.

6. A privacy-preserving computation visualization method based on event-driven sensing, characterized in that, Includes the following steps: The system uses an event-driven image sensor to capture changes in ambient light intensity and outputs a raw event stream. The original event stream is analyzed locally on the sensing device to identify active event areas corresponding to different body parts; Based on the recognition results, a privacy-preserving computational visualization data stream containing body part semantics is generated, wherein the visualization data stream identifies at least one body part through visual features; The privacy-computing visualization data stream is transmitted to a display terminal for human observation; Throughout the entire process, no image frames are generated, stored, or transmitted.

7. The privacy-preserving computational visual sensing unit or method according to any one of claims 1 to 6, characterized in that, In the data processing interface module or method steps, by combining the system's real-time clock information and judging the current scene mode based on the activity intensity of the original event stream, the communication packet parameters in the data transmission process are dynamically adjusted according to different scene modes to achieve adaptive encapsulation and transmission management of the event stream, and to prevent buffer overflow or transmission delay caused by data bursts.

8. A behavior observation system for privacy-sensitive scenarios, characterized in that, include: Multiple privacy-preserving computational vision sensing units, each of which includes an optical lens and an event-driven image sensor, the event-driven image sensor being configured to respond asynchronously to pixel-level illumination changes and output a raw event stream containing spatial coordinates, temporal information, and the polarity of brightness changes, rather than periodic image frames; At least one independent edge computing device is communicatively connected to the plurality of privacy-preserving computational vision sensing units, for receiving the raw event streams output by each sensing unit and performing the following operations locally: The original event stream was analyzed to identify active event regions corresponding to different parts of the human body; A privacy-preserving computation visualization data stream is generated based on the recognition results. The visualization data stream identifies at least one body part through visual features and does not rely on image reconstruction or artificial intelligence recognition. The privacy computing visualization data stream is stored locally and indexed to support data retrieval by time, space, or event type; A display terminal is used to receive and present privacy computing visualization data streams from the edge computing device for human observation; The access control module is used to control access permissions to the raw event stream, the privacy computing visualization data stream, and the stored data. An optional cloud server, communicatively connected to the edge computing device, is used to receive, store, or analyze de-identified privacy computing visualization data streams, or to achieve data synchronization and centralized management across regional systems; In this system, no node generates, stores, or transmits any form of complete image frame, and the output visualization data cannot be restored to image information that can identify an individual, thus achieving privacy protection at the source of data collection.

9. The system according to claim 8, characterized in that, The scenarios include home-based elderly care, family safety monitoring, medical rehabilitation and nursing, or long-term care environments.

10. The privacy-preserving computational vision sensing unit according to any one of claims 1 to 9, characterized in that, The privacy-computing visualization data stream is stored in a computer-readable storage medium or transmitted over a network for use by external devices. The data stream contains visual feature identification information assigned to body parts and does not contain any data that can be used to reconstruct a complete image or identify an individual.

11. The system according to claim 8, characterized in that, The system also includes at least one imaging camera, which is deployed in a public area and connected to the same monitoring network. The video data it outputs is isolated from the privacy computing visualization data stream in terms of storage path, access permissions, and display logic.

Images