Behavior perception system and behavior perception method

Abstract

This invention provides a behavior perception system and method. The system includes a data generation module, a data fusion module, a behavior recognition module, and a result output module. The data generation module acquires biological behavior data and preprocesses it to generate standard behavior data. The data fusion module receives and extracts local features from the standard behavior data, fuses these features to generate a fused feature vector. The behavior recognition module identifies the fused feature vector based on a typical behavior library to output the corresponding behavior category. The behavior recognition module generates behavior attributes based on the standard behavior data and the behavior category. The result output module generates a recognition report based on the behavior category and behavior attributes. This application, through the collaborative work of multiple modules, achieves comprehensive acquisition, deep fusion, and accurate recognition of biological behavior data, solving problems such as single data dimension and poor data fusion effect in existing technologies, and improving the accuracy of behavior recognition.

Description

Technical Field

[0001] This invention relates to the field of environmental behavior perception technology, and in particular to a behavior perception system and behavior perception method. Background Technology

[0002] With the intelligent upgrading of built environments such as smart buildings, ecological farms, and medical and health care facilities, accurate perception of the behavior of organisms (such as humans and animals) is the core foundation for optimizing space management and ensuring biosafety. Currently, the need for behavioral perception of organisms has evolved from simple perception to multi-scenario, high-precision, and low-interference requirements.

[0003] However, existing biological behavior sensing methods are insufficient to meet the application requirements in complex environments. Specifically, most sensing solutions rely on a single sensor, such as a visual sensor or a sound sensor. However, the type and amount of data acquired by a single sensor are limited, and based solely on this data, it is impossible to accurately perceive behavior, especially subtle behaviors, which may go undetected. Some solutions attempt to fuse data from two or more sensors, but these solutions merely involve "simple data stitching" or "mechanical superposition of features," failing to achieve deep fusion and thus limiting the accuracy of behavior recognition.

[0004] Therefore, existing behavior perception schemes suffer from insufficient perception accuracy. Summary of the Invention

[0005] The purpose of this invention is to address the problem of insufficient perception accuracy in existing behavior perception schemes. To solve this problem, an embodiment of this invention discloses a behavior perception system, comprising: a data generation module, which acquires biological behavior data, preprocesses the biological behavior data to generate standard behavior data, and sends the standard behavior data; a data fusion module, which receives the standard behavior data from the data generation module, extracts local features from the standard behavior data, fuses the local features to generate a fused feature vector, and sends the fused feature vector; a behavior recognition module, which stores a typical behavior library, receives the fused feature vector from the data fusion module, performs recognition processing on the fused feature vector according to the typical behavior library to output the corresponding behavior category, and sends the behavior category; furthermore, the behavior recognition module receives the standard behavior data from the data generation module, generates behavior attributes based on the standard behavior data and the behavior category, and sends the behavior attributes; and a result output module, which receives the behavior category and behavior attributes from the behavior recognition module, and generates a recognition report based on the behavior category and behavior attributes.

[0006] According to another specific embodiment of the present invention, the behavior perception system disclosed in this embodiment includes a data generation module comprising a data acquisition unit, a data preprocessing unit, and a data quality monitoring unit; and the biological behavior data includes location data, motion data, and physiological data; wherein the data acquisition unit includes a location data acquisition device for acquiring location data, a motion data acquisition device for acquiring motion data, a physiological data acquisition device for acquiring physiological data, and a supplementary data acquisition device for acquiring supplementary location data and biological movement data; and the data preprocessing unit receives the biological behavior data from the data acquisition unit and preprocesses the biological behavior data to generate standard behavior data; In this process, preprocessing includes at least one of the following: removing abnormal data, data synchronization, and data format standardization. Furthermore, the data quality monitoring unit monitors the signal strength, data loss rate, and power supply status of the location data acquisition unit, motion data acquisition unit, and physiological data acquisition unit. When the signal strength, data loss rate, or power supply status of any acquisition unit is abnormal, a control command is sent to the data acquisition unit to activate a supplementary data acquisition unit. Additionally, the data quality monitoring unit stores and sends abnormal information about the acquisition unit, including the failure time of the location data acquisition unit, motion data acquisition unit, and physiological data acquisition unit, as well as the number of times the supplementary data acquisition unit has been activated.

[0007] According to another specific embodiment of the present invention, the behavior perception system disclosed in this embodiment sends a control command to the data acquisition unit to activate a supplementary data acquisition unit when the signal strength, data packet loss rate, or power supply status corresponding to any acquisition device is abnormal. This includes: when the position data acquisition device is abnormal, activating the supplementary data acquisition unit to obtain supplementary position data and supplementary movement speed data of the organism; and after activating the supplementary data acquisition unit, further including: generating standard behavior data based on the supplementary position data and supplementary movement speed data; when the motion data acquisition device is abnormal, activating the supplementary data acquisition unit to obtain supplementary position data; and after activating the supplementary data acquisition unit, further including: calculating the motion amplitude based on the supplementary position data obtained by the supplementary data acquisition unit, calculating the position change based on the position data, and generating standard behavior data based on the motion amplitude and position change; and when the physiological data acquisition device is abnormal, activating the supplementary data acquisition unit to obtain supplementary position data; and after activating the supplementary data acquisition unit, further including: calculating the motion amplitude based on the supplementary position data obtained by the supplementary data acquisition unit, and inferring standard behavior data based on the motion amplitude and motion data.

[0008] According to another specific embodiment of the present invention, the behavior perception system disclosed in this embodiment of the present invention includes a data fusion module that extracts local features from standard behavior data, including: extracting the slope of the movement trajectory of position data, the rate of change of joint angles of motion data, and the coefficient of variation of heart rate of physiological data; and, when the data preprocessing unit generates standard behavior data, it further includes: generating data confidence scores corresponding to position data, motion data, and physiological data according to the completeness, consistency, and stability of the standard behavior data; the data fusion module performs fusion processing on the local features, including: calculating the initial weights of position data, motion data, and physiological data according to the data confidence scores and using a weighted fusion algorithm, and performing data fusion on the local features according to the initial weights to generate a fused feature vector; wherein, the data fusion module calculates the initial weights according to the following formula: in, For the first The weights of the data types range from 0 to 1, and the initial sum of the weights of the location data, action data, and physiological data is 1. The data confidence level for the Nth type of data is defined; furthermore, the data fusion module also includes a data consistency monitoring unit; wherein, the data consistency monitoring unit generates consistency judgment results for location data, action data, and physiological data, and sends control information to the data generation module to reacquire biological behavior data when the judgment result is that the data is inconsistent.

[0009] According to another specific embodiment of the present invention, the behavior perception system disclosed in this embodiment includes a behavior recognition module that performs recognition processing on the fused feature vector through a neural network classification model and outputs a behavior category; furthermore, the behavior recognition module also outputs a category confidence level corresponding to the behavior category; and the behavior attributes include the behavior occurrence time, the behavior occurrence location, the behavior duration, and the behavior frequency; the behavior recognition module also includes an anomaly warning unit; wherein the anomaly warning unit stores preset abnormal behavior standards, and the anomaly warning unit generates graded warning information based on the behavior category, behavior attributes, and abnormal behavior standards; and the recognition report generated by the result output module includes at least one of a structured behavior report, a two-dimensional heatmap, and a three-dimensional behavior trajectory animation; and the result output module also includes an external protocol port, through which the result output module sends the recognition report to the target device.

[0010] According to another specific embodiment of the present invention, the behavior perception system disclosed in this embodiment further includes a data security module, which includes an encryption management unit, a permission setting unit, a behavior recording unit, a storage unit, and a privacy protection unit. The encryption management unit is connected to the data generation module, the data fusion module, the behavior recognition module, and the result output module, and performs encryption and decryption processing on the data when the data generation module, the data fusion module, the behavior recognition module, and the result output module send and receive data. The permission setting unit sets operation permissions for all data, including permissions for administrators only, permissions for administrators and operators only, and permissions for visitors only. The behavior recording unit records all operations corresponding to each piece of data and generates a behavior log, which is stored in a read-only partition of the storage unit. The privacy protection unit desensitizes the action data in the standard behavior data and anonymizes the standard behavior data corresponding to each organism, generating an identification code for each organism, and stores the identification code and the corresponding standard behavior data in the storage unit.

[0011] According to another specific embodiment of the present invention, the behavior perception system disclosed in this embodiment further includes a feedback optimization module, which includes a touch interaction unit and / or a voice interaction unit, and an optimization unit; wherein, the touch interaction unit receives externally input text interaction information and / or gesture interaction information; the voice interaction unit receives externally input voice interaction information; and the optimization unit extracts feedback information based on the text interaction information, gesture interaction information, and / or voice interaction information, and optimizes the neural network classification model of the behavior recognition module based on the feedback information.

[0012] According to another specific embodiment of the present invention, the behavior perception system disclosed in this embodiment further includes a data correction module, which comprises an environmental data acquisition unit and a correction unit; wherein, the environmental data acquisition unit acquires temperature data, humidity data, light data, and sound data of the target area, and the environmental data acquisition unit includes a temperature and humidity sensor for acquiring temperature and humidity data, a light sensor for acquiring light data, and a noise sensor for acquiring sound data; the correction unit corrects the position information and / or action information in the standard behavior data based on the temperature and humidity data and using the following formula:

[0013]

[0014] in, For the corrected location information and / or motion information; Raw data acquired by the temperature and humidity sensor; The temperature influence coefficient ranges from 0.005 to 0.01, indicating that the correction range for the data is 0.5% to 1% for every 1°C deviation of the temperature from the standard value; T represents the temperature data in °C. The standard temperature is 25℃. The humidity influence coefficient ranges from 0.003 to 0.008, indicating that the correction margin for the data is 0.3% to 0.8% for every 1% RH deviation of the humidity from the standard value. Humidity data, in %RH; The standard humidity is set at 50%RH. The correction unit corrects the exposure parameters of the motion data acquisition unit in the data generation module based on the illumination data. Specifically, when the illumination data is less than or equal to the preset illumination threshold, the exposure of the motion data acquisition unit is increased. The correction unit corrects the voice interaction unit of the feedback optimization module based on the sound data. Specifically, when the sound data is greater than the preset sound threshold, the input data of the voice interaction unit is stopped.

[0015] This invention discloses a behavior perception method applicable to the behavior perception system described in any of the above embodiments. The behavior perception method includes: acquiring biological behavior data and preprocessing the biological behavior data to generate standard behavior data; extracting local features from the standard behavior data and fusing the local features to generate a fused feature vector; identifying the fused feature vector based on a typical behavior library and using a neural network classification model, and outputting the corresponding behavior category; generating behavior attributes based on the standard behavior data and the behavior category; and generating an identification report based on the behavior category and behavior attributes.

[0016] According to another specific embodiment of the present invention, the behavior perception method disclosed in this embodiment further includes: encrypting biological behavior data, standard behavior data, fused feature vectors, behavior categories, behavior attributes, and recognition reports, and setting operation permissions; and / or generating behavior logs based on all operations corresponding to the biological behavior data, standard behavior data, fused feature vectors, behavior categories, behavior attributes, and recognition reports; and / or acquiring externally input text interaction information, gesture interaction information, and / or voice interaction information, extracting feedback information based on the text interaction information, gesture interaction information, and / or voice interaction information, and optimizing the neural network classification model based on the feedback information; and / or acquiring environmental data of the target area, and correcting the standard behavior data and voice interaction information based on the environmental data.

[0017] The beneficial effects of this invention are:

[0018] The behavior perception system and method provided in this application achieve comprehensive collection, deep fusion and accurate recognition of biological behavior data through the collaborative work of multiple modules. It solves the problems of single data dimension and poor data fusion effect in the prior art and has the advantage of improving the accuracy of behavior recognition. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of the behavior perception system provided in an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of the data generation module in the behavior perception system provided in an embodiment of the present invention;

[0021] Figure 3 This is another structural schematic diagram of the behavior perception system provided in an embodiment of the present invention;

[0022] Figure 4 This is a graph showing the comparison of behavior recognition accuracy before and after data fusion in the behavior perception system provided in this embodiment of the invention;

[0023] Figure 5 This is a diagram showing the effect of the behavior perception system provided in this embodiment of the invention after data correction under different environmental parameters;

[0024] Figure 6 This is a flowchart illustrating the behavior perception method provided in an embodiment of the present invention. Detailed Implementation

[0025] Example 1:

[0026] As described in the background section, existing behavior perception schemes suffer from insufficient perception accuracy. To address these technical problems, this embodiment provides a behavior perception system. (Reference) Figure 1 The behavior perception system includes a data generation module, a data fusion module, a behavior recognition module, and a result output module. The data generation module acquires biological behavior data, preprocesses it to generate standard behavior data, and then sends the standard behavior data. The data fusion module receives the standard behavior data from the data generation module, extracts local features from the standard behavior data, fuses these local features to generate a fused feature vector, and then sends the fused feature vector. The behavior recognition module stores a typical behavior library. It receives the fused feature vector from the data fusion module, performs recognition processing on the fused feature vector according to the typical behavior library to output the corresponding behavior category, and then sends the behavior category. Furthermore, the behavior recognition module receives the standard behavior data from the data generation module, generates behavior attributes based on the standard behavior data and the behavior category, and then sends the behavior attributes. The result output module receives the behavior category and behavior attributes from the behavior recognition module, and generates a recognition report based on the behavior category and behavior attributes.

[0027] Specifically, the data generation module can be implemented through a combination of various sensors. For example, it can acquire motion data using cameras or inertial measurement units, location data using cameras, GPS, or UWB positioning modules, and physiological data using wearable devices or heart rate sensors. Biological behavior data refers to data related to biological behavior directly acquired through the aforementioned sensors. The preprocessing process can include data cleaning, time alignment, and normalization. Data cleaning can employ statistical outlier removal methods, time alignment can use dynamic time warping algorithms, and normalization can use minimum-maximum standardization methods. Standard behavior data refers to the preprocessed biological behavior data. The data fusion module can employ feature-level fusion methods, such as principal component analysis or independent component analysis to reduce and fuse multi-source features. Specifically, the data fusion module can be implemented using a weighted fusion algorithm combined with a CNN-LSTM deep learning model. The CNN-LSTM deep learning model has a 128-dimensional input layer, two hidden layers with 64 neurons each, and a 20-dimensional output layer. The behavior recognition module can employ classification algorithms such as Support Vector Machines, Random Forests, or Deep Neural Networks. A typical behavior library can be trained using historical data. Behavior categories include specific behaviors such as "walking," "running," "falling," or "specific gestures." Behavior attributes are quantitative descriptions of these different behavior categories, such as walking time, running speed, frequency of falls, and the amplitude of gestures. The output module generates a recognition report that can be a structured report in JSON format, or it can generate a two-dimensional heatmap and a three-dimensional trajectory animation using visualization tools.

[0028] This structure, through the collaborative work of multiple modules, enables precise perception of biological behavior. The data generation module is responsible for collecting and standardizing raw data, providing high-quality input for subsequent data processing. The data fusion module, through feature extraction and fusion, addresses the limitation of data volume from a single sensor, enhancing feature representation capabilities. The behavior recognition module, combining a typical behavior library and a neural network model, achieves accurate classification and recognition of complex behaviors. The results output module presents the recognition results. Furthermore, the behavior recognition module not only identifies behavior categories but also generates behavioral attributes that quantify and describe behaviors based on standard behavior data, enriching the content of the recognition results and thus improving the accuracy and scope of recognition. Moreover, the results output module generates richer and more valuable recognition reports based on behavior categories and attributes, meeting the needs for behavioral information in different application scenarios.

[0029] Furthermore, in this behavior perception system, reference Figure 2The data generation module includes a data acquisition unit, a data preprocessing unit, and a data quality monitoring unit. Furthermore, the biological behavior data includes location data, motion data, and physiological data. The data acquisition unit includes a location data acquisition unit for acquiring location data, a motion data acquisition unit for acquiring motion data, a physiological data acquisition unit for acquiring physiological data, and a supplementary data acquisition unit for acquiring supplementary location data and biological movement data. The location data acquisition unit can be implemented using 8 to 12 infrared sensors. An infrared transmitter emits infrared light, which is received by an infrared receiver after being reflected back by the organism. Distance coordinate transformation is then performed to calculate the organism's specific location. The motion data acquisition unit can be implemented using 4 to 6 4K high-definition cameras. The cameras acquire images or videos of biological behavior, and the relevant motion data is obtained through image or video recognition processing. The physiological data acquisition unit can be implemented using wearable devices to monitor data such as the organism's heart rate, blood oxygen saturation, and cadence. Supplemental data acquisition can be achieved through backup infrared sensors, cameras, and wearable devices, or through 2 to 4 channels of 24 GHz millimeter-wave radar, with a ranging accuracy of ±0.1 m and a velocity measurement range of 0.1 m / s to 10 m / s. Millimeter-wave radar can measure the distance between the radar and the organism in real time and accurately, and then accurately calculate the speed and direction of the organism's movement based on this distance. (Reference) Figure 4 Extensive research data shows that, due to their inherent limitations, the accuracy of behavior recognition using a single sensor is generally below 75%. However, this application significantly improves the accuracy of behavior recognition by deeply fusing data from multiple sensors.

[0030] Preferably, the infrared sensor has a detection range of 3m to 8m and a detection angle of 120° to 150° to more comprehensively sense organisms within the predetermined range and prevent omissions. The 4K high-definition camera has a frame rate of 25fps to 30fps and supports a wide dynamic range, enabling continuous acquisition of organism movements. In practical applications, all processing modules are integrated onto a single circuit board, allowing wearable devices to connect via Bluetooth. The data acquisition unit supports dual transmission modes: Ethernet, LoRa, and WiFi 6. The Ethernet bandwidth is gigabit-level with a transmission latency of no more than 10ms; the LoRa transmission distance is 1km to 3km with a power consumption of no more than 50μA; and the WiFi 6 transmission rate is no less than 1.2Gbps with a concurrent connection count of no less than 100. Furthermore, when using this data acquisition unit for sampling, the data sampling frequency can be dynamically adjusted according to the scenario. For example, in static scenarios such as indoor offices, the sampling frequency is set to 1Hz to 3Hz, corresponding to data collection every 333ms to 1000ms. In dynamic scenarios such as shopping mall pedestrian traffic, the sampling frequency is set to 5Hz to 10Hz, corresponding to data collection every 100ms to 200ms. In other preferred implementations, when a sensor malfunction occurs, the sampling frequency for the corresponding scenario can be temporarily increased by 20% to 30%. For example, when an infrared sensor malfunctions, the sampling frequency of the camera and wearable device can be increased.

[0031] Furthermore, in this behavior perception system, the data preprocessing unit receives biological behavior data from the data acquisition unit and preprocesses the biological behavior data to generate standard behavior data. Preprocessing includes at least one of the following: outlier removal, data synchronization, and data format standardization. Specifically, outlier removal can be achieved by setting a threshold range or using a statistical outlier detection algorithm. Preferably, the 3σ principle is used for outlier removal. In specific implementation, the mean μ and standard deviation σ of the single-sensor data are first calculated, and then noisy data exceeding the range of μ minus 3σ to μ plus 3σ is removed. Data synchronization can be achieved using timestamp alignment or interpolation compensation methods. Preferably, multi-source data synchronization can be achieved based on GPS timestamps, where the GPS timestamp accuracy does not exceed 1ms to ensure that the time deviation between the location data collected by the infrared sensor, the motion data collected by the camera, and the physiological data collected by the wearable device does not exceed 10ms. Data format standardization can be achieved by unifying the sampling frequency or data normalization processing. Preferably, the data format standardization employs a min-max algorithm, mapping the organism's location coordinates to a range of 0 to 100 to correspond to the built-up environment plane coordinates, mapping limb movement amplitude to a range of 0 to 1 to correspond to a movement range of 0° to 180°, and mapping physiological data such as heart rate to a range of 0 to 1 to correspond to a range of 40 beats / minute to 180 beats / minute. After preprocessing, a preprocessed dataset in JSON format can be output, where each data entry includes the sensor ID, acquisition timestamp, and standardized data value.

[0032] Furthermore, in this behavior perception system, the data quality monitoring unit monitors the signal strength, data packet loss rate, and power supply status of the location data acquisition unit, motion data acquisition unit, and physiological data acquisition unit. When any acquisition unit experiences an abnormal signal strength, data packet loss rate, or power supply status, the unit sends a control command to the data acquisition unit to activate a supplementary data acquisition unit. The data quality monitoring unit also stores and sends abnormal information about the acquisition unit's malfunctions, including the failure time of the location data acquisition unit, motion data acquisition unit, and physiological data acquisition unit, as well as the number of times the supplementary data acquisition unit has been activated. Specifically, the signal strength monitoring of the data quality monitoring unit can be achieved through RSSI value detection. In a preferred implementation, the data quality monitoring unit is implemented using a TICC2650 wireless chip, which can detect the signal strength, data packet loss rate, and wearable device battery level of each sensor in real time. The specific signal strength detection standard is that the WiFi signal strength is not lower than -85dBm, and the LoRa signal strength is not lower than -110dBm. When the signal strength is lower than the corresponding threshold, a signal enhancement command can also be triggered. The data packet loss rate can be achieved through data packet sequence number verification or quantity comparison. In a preferred implementation, the data packet loss rate is calculated by comparing the number of sent data packets with the number of received data packets. Furthermore, a data packet loss rate of no more than 3% can be set as the normal state; a retransmission mechanism is triggered when the data packet loss rate is between 3% and 5%; and an alarm is triggered when the data packet loss rate exceeds 5%. Power supply status can be monitored using voltage and current sensors. The camera and infrared sensor can be continuously powered by a constant power supply or powered by an external battery. Wearable devices generally use battery power. When powered by battery, the battery level of each signal collector is acquired at predetermined intervals, and a battery level of no less than 20% is set as the normal state. A low battery reminder is sent to the user terminal when the battery level is between 10% and 20%, and the device is marked as offline when the battery level is below 10%. The startup control of the supplementary data collector can be achieved through a hardware switching circuit or a software enable signal.

[0033] This approach, through the redundant configuration of the data quality monitoring unit and supplementary data acquisition units, avoids data loss caused by the failure of a single sensor. When the main acquisition unit experiences signal abnormalities, the backup acquisition unit is automatically activated to supplement data or the sampling frequency of other acquisition units is increased, ensuring the continuity of behavioral data. Simultaneously, by recording anomaly information and activation counts, data support is provided for subsequent equipment maintenance.

[0034] Furthermore, in this behavior perception system, when the signal strength, data packet loss rate, or power supply status corresponding to any data collector is abnormal, a control command is sent to the data acquisition unit to activate a supplementary data collector. This includes: when the location data collector malfunctions, activating the supplementary data collector to obtain supplementary location data and supplementary movement speed data of the organism; and after activating the supplementary data collector, generating standard behavior data based on the supplementary location data and supplementary movement speed data. For example, when the infrared sensor fails, millimeter-wave radar can be activated to supplement the location and movement speed data.

[0035] When the motion data acquisition device malfunctions, a supplementary data acquisition device is activated to obtain supplementary position data. Furthermore, after activating the supplementary data acquisition device, the process includes: calculating the motion amplitude based on the supplementary position data, calculating the position change based on the position data, and generating standard behavior data based on the motion amplitude and position change. For example, when the camera malfunctions, an infrared sensor and millimeter-wave radar can be activated to jointly determine the behavior based on position change and motion amplitude.

[0036] When the physiological data acquisition device malfunctions, a supplementary data acquisition device is activated to obtain supplementary location data. Furthermore, after activating the supplementary data acquisition device, the process includes: calculating the amplitude of movement based on the supplementary location data acquired by the supplementary data acquisition device, and inferring standard behavioral data based on the amplitude of movement and the movement data. For example, when the wearable device is offline, the camera can be activated to indirectly infer physiological state-related behaviors through facial expressions and amplitude of movement; for example, inferring "rapid running" behavior from "rapid breathing."

[0037] Of course, supplementary data acquisition units can be implemented in various other ways. For example, a supplementary data acquisition unit can be an inertial measurement unit (IMU) that acquires supplementary position and velocity data via accelerometers and gyroscopes. Furthermore, a supplementary data acquisition unit can also be an infrared sensor or an ultrasonic sensor, used to acquire supplementary position data when the motion data acquisition unit malfunctions. As a preferred implementation, the supplementary data acquisition unit can integrate multiple sensors, selectively activating the corresponding sensors based on different abnormal situations. Thus, the supplementary data acquisition unit can flexibly respond to abnormal situations of different types of acquisition units, ensuring the integrity of standard behavioral data.

[0038] This approach, by dynamically activating supplementary data collectors and employing differentiated data processing strategies based on the type of malfunctioning collector, effectively solves the problem of data loss caused by single sensor anomalies. Specifically, for location data collector anomalies, standard behavioral data is directly generated using supplementary data; for motion data collector anomalies, motion amplitude is indirectly calculated using supplementary location data; and for physiological data collector anomalies, inference is performed by combining motion amplitude and motion data. This allows the behavior perception system to ensure continuity and reliability by adaptively adjusting the data processing logic in the event of different collector anomalies.

[0039] Furthermore, in this behavior perception system, the data fusion module extracts local features from standard behavior data, including: the slope of the movement trajectory from position data, the rate of change of joint angles from action data, and the coefficient of variation of heart rate from physiological data. Specifically, the slope of the movement trajectory can be obtained by calculating the rate of change of displacement between adjacent sampling points using the difference method; the rate of change of joint angles can be obtained by integrating and differentiating the angular velocity data collected by the inertial measurement unit; and the coefficient of variation of heart rate can be calculated by the standard deviation of the peak interval of the electrocardiogram signal.

[0040] Furthermore, in this behavior perception system, when the data preprocessing unit generates standard behavior data, it also includes: generating data confidence scores for location data, action data, and physiological data based on the completeness, consistency, and stability of the standard behavior data. The generation of data confidence scores can employ statistical distribution-based methods, such as quantifying completeness by calculating data missing rate, noise ratio, or temporal continuity indicators; quantifying consistency by cross-validating the spatiotemporal alignment of data from different sensors; and quantifying stability through variance analysis within a sliding window.

[0041] Furthermore, in this behavior perception system, the data fusion module performs fusion processing on local features, including: calculating initial weights for location data, action data, and physiological data respectively based on data confidence and using a weighted fusion algorithm; and fusing local features according to the initial weights to generate a fused feature vector. Specifically, the weighted fusion algorithm can employ a dynamic adjustment strategy, automatically reducing the weight ratio of a certain type of data when its confidence level falls below a threshold. The data fusion module calculates the initial weights according to the following formula:

[0042]

[0043] in, For the first The weights of the data types range from 0 to 1, and the initial sum of the weights of the location data, action data, and physiological data is 1. This represents the data confidence score for category X. Specifically, the data confidence score is calculated based on the data's completeness, consistency, and stability. Specifically, completeness is assessed based on missing data, and a completeness confidence score is calculated accordingly. Consistency is assessed based on the data's matching degree with other data, and a consistency confidence score is calculated accordingly. Stability is assessed based on the fluctuations of data across N consecutive sampling periods, and a stability confidence score is calculated accordingly. Finally, the data confidence score is calculated by comprehensively considering the completeness confidence score, consistency confidence score, and stability confidence score. For example, the integrity confidence level is 1 when there are no missing values, 0.8 when one data point is missing, and 0.5 when two or more data points are missing. Similarly, the consistency confidence level is 1 when the data matches at least 90% of other data sources, 0.7 when the match is between 70% and 90%, and 0.3 when the match is below 70%. The stability confidence level is 1 when the fluctuation of data collected over five consecutive data collections does not exceed 5%, 0.8 when the fluctuation is between 5% and 10%, and 0.5 when the fluctuation is above 10%. The data confidence level is calculated using the formula: Data Confidence Level = Integrity Confidence Level * Consistency Confidence Level * Stability Confidence Level.

[0044] Furthermore, in this behavior perception system, the data fusion module also includes a data consistency monitoring unit. This unit generates consistency judgment results for location data, motion data, and physiological data. When the judgment result indicates data inconsistency, it sends control information to the data generation module to reacquire biological behavior data. Data consistency monitoring can be achieved by constructing a multi-source data association matrix. When the matrix feature values ​​deviate from a preset range, an inconsistency alarm is triggered. For example, consistency judgment could be determined when location data shows the human body is stationary but motion data shows significant movement, indicating data inconsistency. If inconsistency is found, a secondary acquisition verification is triggered, i.e., data is reacquired at intervals of 50ms to 100ms. Preferably, during data preprocessing and consistency judgment, data features with a confidence level of not less than 0.7 to 0.85 are preferentially retained.

[0045] In this way, data confidence can dynamically reflect the reliability differences of data from different sensors, and combined with weight allocation, adaptive fusion at the feature level can be achieved, thereby improving the accuracy of behavior recognition. The data consistency monitoring unit can detect data deviations caused by sensor anomalies or environmental interference in a timely manner before fusion, preventing erroneous features from entering the recognition process.

[0046] Furthermore, in this behavior perception system, typical behaviors in the typical behavior library include, but are not limited to, walking at a constant speed, running quickly, sitting for extended periods, standing, climbing stairs, picking up and placing objects, bending over to pick up objects, leaning, talking, and writing. Different behaviors can be predefined; for example, sitting for extended periods is defined as a static state lasting at least 30 minutes, picking up and placing objects is defined as raising the hand at a height of at least 50cm, and bending over to pick up objects is defined as an action where the angle between the torso and the ground does not exceed 45°. Typical animal behaviors include walking at a constant speed, running quickly, lying down, eating, and drinking. The behavior recognition module uses a neural network classification model to identify and process the fused feature vectors and output the behavior category. The neural network classification model can be implemented using a convolutional neural network or a recurrent neural network. Convolutional neural networks are suitable for processing fused feature vectors with obvious spatial features, while recurrent neural networks are suitable for processing fused feature vectors with obvious time-series features. Preferably, the neural network classification model uses a trained ResNet-18 classification model to identify the fused feature vectors. The ResNet classification model has a training sample size of at least 50,000 samples, covering people of different ages, heights, and movement habits. The behavior recognition module also outputs a category confidence score corresponding to the behavior category. The category confidence score is calculated using the softmax function, with a value ranging from 0 to 1; a higher value indicates higher reliability of the classification result. Preferably, the category confidence score is accurate to two decimal places, the recognition latency is no more than 500ms (i.e., the time from receiving the fused features to outputting the recognition result is no more than 500ms), the accuracy rate for 100 consecutive recognitions is no less than 90%, and it supports behavior duration statistics, such as sitting duration accurate to 1 minute.

[0047] Furthermore, in this behavior perception system, behavior attributes include the time of occurrence of the behavior, the location of occurrence of the behavior, the duration of the behavior, and the frequency of the behavior.

[0048] Furthermore, in this behavior perception system, the behavior recognition module also includes an anomaly warning unit. This unit stores preset abnormal behavior standards and generates tiered warning information based on behavior category, behavior attribute, and the abnormal behavior standards. The abnormal behavior standards can be set to different threshold ranges. For example, in a medical scenario, abnormal heart rate and falls can be set as Level 1 warnings, while prolonged stillness can be set as Level 2 warnings. The abnormal behavior standards need to be pre-defined. For instance, several different abnormal behaviors can be set, with some categories classified as Level 1 warnings and others as Level 2 warnings. For example, the first category is sitting for more than 2 hours, defined as continuous recognition of "sitting for a long time"; the second category is a sudden fall, defined as a sudden increase in the amplitude of movement exceeding 45° with the amplitude not exceeding 5° within the following 10 seconds; the third category is a sudden increase in heart rate exceeding 30 beats / minute, defined as a heart rate rising from a resting value of 60-80 beats / minute to at least 110 beats / minute within 1 minute; and the fourth category is prolonged stillness, defined as no movement for more than 120 minutes.

[0049] Preferably, the output module of the behavior perception system is also connected to the administrator terminal to send abnormal information. This abnormal information includes the type of abnormal behavior, its location, a real-time data screenshot, and the warning level. The warning can be in the form of a buzzer or a light.

[0050] Furthermore, in this behavior perception system, the recognition report generated by the result output module includes at least one of the following: a structured behavior report, a two-dimensional heatmap, and a three-dimensional behavior trajectory animation. The result output module also includes an external protocol port, through which the recognition report is sent to the target device. The external protocol port supports at least one of HTTP, MQTT, or WebSocket protocols, configured according to the communication requirements of the target device. The target device can be, for example, a PC management terminal or a tablet terminal. The PC management terminal is compatible with Windows and Linux systems, while the tablet terminal is compatible with Android and iOS systems. The structured behavior report includes the behavior type, occurrence time, occurrence location, behavior duration, and frequency statistics. The two-dimensional heatmap can specifically use Jet color mapping, for example, using red to represent areas with a behavior frequency of at least 10 times / hour, and blue to represent areas with a behavior frequency of no more than 2 times / hour, and users can switch between different time periods to view high-incidence areas. The three-dimensional behavior trajectory animation is generated based on the Unity engine, supports viewpoint rotation and zoom operations, and its trajectory line color corresponds to the behavior type.

[0051] Furthermore, this behavior perception system also includes a data security module, which comprises an encryption management unit, a permission setting unit, a behavior recording unit, a storage unit, and a privacy protection unit. The encryption management unit is connected to the data generation module, data fusion module, behavior recognition module, and result output module, and performs encryption and decryption processing on the data during data transmission and reception in these modules. The permission setting unit sets operation permissions for all data, including administrator-only, administrator and operator-only, and visitor-only permissions. The behavior recording unit records all operations corresponding to each piece of data and generates a behavior log, which is stored in a read-only partition of the storage unit. The privacy protection unit desensitizes the action data in the standard behavior data and anonymizes the standard behavior data corresponding to each organism, generating an identification code for each organism. The identification code and the corresponding standard behavior data are stored in the storage unit.

[0052] The encryption management unit can use AES-256 or RSA algorithms for data encryption and decryption, with key management protected by a hardware security module. The permission setting unit implements hierarchical permissions through a role-based access control model, specifically through a fine-grained control achieved by mapping user groups to permission tags. Preferably, three levels of access permissions can be set: administrator permissions, operator permissions, and visitor permissions. Administrators can modify system parameters and view all data; operators can only view and export data from authorized areas; and visitors can only view public behavior statistics reports. Permission allocation is implemented through a role-based access control (RBAC) mechanism. The behavior recording unit uses blockchain technology to ensure the immutability of logs, where operation records include timestamps, operation types, and operator identity information. Preferably, logs can include operator ID, access time, data type, operation content, and IP address. Operation content includes querying, exporting, and modifying; IP addresses include public and private IP addresses. Furthermore, logs are preferably stored in a read-only partition with a retention period of no less than 6 months, and can be exported in CSV or PDF format for audit analysis. The read-only partitions of the storage unit are implemented through write protection mechanisms, such as a read-only file system or a physical write protection switch. Specifically, the storage unit uses a MongoDB distributed database, which supports 3-replica disaster recovery backups with a backup cycle of 1 hour and backup data retention for 7 days. The database is deployed on a private server to avoid public network exposure. The privacy protection unit's anonymization processing can employ data masking or generalization techniques, or use the OpenCV image processing library to anonymize the collected image data. Anonymization is achieved by generating irreversible identification codes using hash algorithms or UUIDs, such as the SHA-256 algorithm. Preferably, when the behavior perception system is used for human behavior perception, the privacy protection unit uses the OpenCV image processing library to anonymize the collected human image data, blurring facial features using a Gaussian blur algorithm, including the eyes, nose, and mouth, and simultaneously using a mosaic algorithm to hide identifiable markings on clothing, such as logos and text. Meanwhile, the privacy protection unit stores physiological data anonymously and uses UUID (Universally Unique Identifier) ​​to generate a unique user ID, replacing real name, ID number and other identity information. The user ID and real identity information are set to be queried only by the administrator through an encrypted database.

[0053] This approach, through an encrypted management unit ensuring the security of data transmission and storage, an access control unit enabling controlled data access, a behavior recording unit providing operational traceability, and a privacy protection unit ensuring the protection of biometric privacy information, forms a complete security system that effectively addresses the risks of data leakage and misuse in behavior perception systems.

[0054] Furthermore, this behavior perception system also includes a feedback optimization module, which comprises a touch interaction unit and / or a voice interaction unit, as well as an optimization unit. The touch interaction unit receives externally input text interaction information and / or gesture interaction information; the voice interaction unit receives externally input voice interaction information; and the optimization unit extracts feedback information based on the text interaction information, gesture interaction information, and / or voice interaction information, and optimizes the neural network classification model of the behavior recognition module based on the feedback information.

[0055] Specifically, the touch interaction unit can use a capacitive or infrared touchscreen to enable text or gesture input, and interpret user gesture commands through a gesture recognition algorithm. The voice interaction unit can integrate a microphone array and a voice recognition chip, supporting both offline voice command recognition and online voice interaction modes. The optimization unit extracts keywords and semantic features from the interaction information using natural language processing technology to form structured feedback data, and uses an incremental learning algorithm to adjust the parameters of the neural network classification model. Incremental learning can be implemented using online gradient descent or mini-batch stochastic gradient descent, calculating the loss function only for new feedback data and backpropagating it during each model update. In practical implementation, only the touch interaction unit and optimization unit can be set up to meet the needs of quiet environments; only the voice interaction unit and optimization unit can be set up to improve the convenience of optimized information input; or both can be set up simultaneously to improve the accuracy of optimized information input.

[0056] Preferably, the feedback information is the user's corrective feedback on the recognition result, and the feedback types include, but are not limited to, recognition error, category bias, and low category confidence. Among them, recognition error means that the actual behavior does not match the output category at all; category bias means that the behavior category is similar but not accurate, such as recognizing "walking slowly" as "standing"; low confidence means that the category confidence is lower than, for example, a threshold of 0.7, but the recognition result is correct.

[0057] Preferably, the neural network classification model of the behavior recognition module is optimized based on the feedback information. Specifically, the feedback information can be labeled as training samples, which are labeled with correct behavior information, such as the correct behavior category, feedback type, and user ID. The neural network classification model of the behavior recognition module is then updated through incremental learning.

[0058] Preferably, the optimization unit also records the frequency of the feedback information, and when the frequency reaches a threshold, such as 5 times, it triggers a model-focused optimization mechanism, such as increasing the weight of the training samples for that behavior category, retraining the local parameters of the model, and generating and pushing optimization suggestions, such as suggesting different angle samples for the corresponding behavior.

[0059] This approach, employing both touch and voice interaction, improves the convenience and completeness of feedback information collection. Furthermore, incremental learning enables continuous model optimization, avoiding the resource consumption associated with retraining the entire model. In addition, the feedback mechanism forms a closed loop with the behavior recognition module, allowing behavior recognition accuracy to continuously improve with usage.

[0060] Furthermore, this behavior perception system also includes a data correction module, which comprises an environmental data acquisition unit and a correction unit. The environmental data acquisition unit acquires temperature, humidity, light, and sound data of the target area. This unit includes a temperature and humidity sensor for acquiring temperature and humidity data, a light sensor for acquiring light data, and a noise sensor for acquiring sound data. Specifically, the temperature and humidity sensor is an SHT30 sensor, with a measurement range of -40℃ to 125℃ and 0%RH to 100%RH, with a temperature measurement accuracy of ±0.2℃ and a humidity measurement accuracy of ±2%RH. The light sensor is a BH1750 light sensor, with a measurement range of 0lx to 65535lx and a measurement accuracy of ±1lx. The noise sensor is an A402 noise sensor, with a measurement range of 30dB to 130dB and a measurement accuracy of ±1dB.

[0061] The correction unit corrects the position and / or motion information in the standard behavior data based on temperature and humidity data, and using the following formula:

[0062]

[0063] in, For the corrected location information and / or motion information; Raw data acquired by the temperature and humidity sensor; The temperature influence coefficient ranges from 0.005 to 0.01, indicating that the correction range for the data is 0.5% to 1% for every 1°C deviation of the temperature from the standard value; T represents the temperature data in °C. The standard temperature is 25℃. The humidity influence coefficient ranges from 0.003 to 0.008, indicating that the correction margin for the data is 0.3% to 0.8% for every 1% RH deviation of the humidity from the standard value. Humidity data, in %RH; The standard humidity is 50%RH.

[0064] The correction unit adjusts the exposure parameters of the motion data acquisition unit in the data generation module based on illumination data. Specifically, when the illumination data is less than or equal to a preset illumination threshold, the exposure of the motion data acquisition unit is increased. For example, when the illumination intensity L is below 500 lx, the camera exposure is automatically increased by 20% to 50%. The correction unit also adjusts the voice interaction unit of the feedback optimization module based on sound data. Specifically, when the sound data exceeds a preset sound threshold, input data reception from the voice interaction unit is stopped. For example, when the noise level N is above 80 dB, input data reception from the voice interaction unit is stopped.

[0065] This approach allows for the elimination of position measurement errors caused by environmental factors through temperature and humidity compensation; ensures motion capture quality under low-light conditions through adaptive lighting adjustment; and improves the reliability of voice interaction through intelligent noise filtering. (Reference) Figure 5 After correction by the data correction module, the deviation of the original data under high temperature, high humidity and high noise environments is significantly reduced.

[0066] Example 2:

[0067] Based on the behavior perception system provided in Embodiment 1, this embodiment also provides a behavior perception method applicable to the behavior perception system as described in Embodiment 1.

[0068] Specifically, refer to Figure 6 The behavior perception method includes the following steps:

[0069] Acquire biological behavior data and preprocess the biological behavior data to generate standard behavior data.

[0070] Local features are extracted from standard behavioral data, and these local features are then fused to generate a fused feature vector.

[0071] Based on a typical behavior library, a neural network classification model is used to identify and fuse feature vectors and output the corresponding behavior categories. Behavioral attributes are generated based on standard behavior data and behavior categories.

[0072] An identification report is generated based on the behavior category and behavior attributes.

[0073] By employing these steps, this behavior perception method addresses the limitations of existing technologies, such as the limited data volume from a single sensor and insufficient depth of feature fusion, through multi-source data acquisition, deep feature fusion, and intelligent behavior recognition. Specifically, data preprocessing and confidence-based feature fusion improve data quality and the reliability of feature fusion, while the neural network classification model further enhances the accuracy of behavior recognition.

[0074] Furthermore, this behavior perception method also includes: encrypting biological behavior data, standard behavior data, fused feature vectors, behavior categories, behavior attributes, and recognition reports, and setting operation permissions.

[0075] Furthermore, this behavior perception method also includes generating behavior logs based on all operations corresponding to biological behavior data, standard behavior data, fused feature vectors, behavior categories, behavior attributes, and recognition reports.

[0076] Furthermore, this behavior perception method also includes: acquiring external input text interaction information, gesture interaction information, and / or voice interaction information, extracting feedback information based on the text interaction information, gesture interaction information, and / or voice interaction information, and optimizing the neural network classification model based on the feedback information.

[0077] Furthermore, this behavior perception method also includes: acquiring environmental data of the target area, and correcting standard behavior data and voice interaction information based on the environmental data.

[0078] It should be noted that the specific data processing details in the behavior perception method provided in this embodiment have been described in detail in Embodiment 1, and will not be repeated in this embodiment.

[0079] Next, we will use the application of this sensing method in a specific scenario of a nursing home as an example. Suppose that the sensing system is applied in a medium-sized nursing home in a city, with 5 floors, 12 rooms, 2 common activity areas, and 1 dining room. It needs to monitor the daily behaviors of the elderly (such as prolonged sitting, climbing stairs, and falls) to ensure safety and optimize care services. The system will be equipped with 8 infrared sensors, 4 4K high-definition cameras, 2 24GHz millimeter-wave radars, and 20 wearable devices.

[0080] When implementing the sensing method, the data generation module is deployed first. Infrared sensors are installed at the entrances of each room and stairwells, 4K high-definition cameras are installed in the public activity areas and the dining room, and 24GHz millimeter-wave radars are installed at stair corners to monitor movements up and down stairs. Bluetooth wearable devices are provided for the elderly to collect heart rate and cadence data. Data transmission uses both Ethernet and LoRa modes. Ethernet connects the cameras and radars; LoRa connects the infrared sensors and wearable devices. Simultaneously, SHT30 temperature and humidity sensors are deployed to detect temperature and humidity in the dining room; BH1750 light sensors are deployed to detect light in the public activity areas; and A402 noise sensors are deployed to detect noise in the dining room. T and H are then substituted into the compensation formula. ,in =0.008、 =25℃ =0.005、 =50%RH. Assume the initial detection range of a certain infrared sensor is... =4m, calculated to =4×[1-0.008×(28-25)-0.005×(65-50)]=4×(1-0.024-0.075)=4×0.901=3.604m, which is consistent with the actual distance after correction. Assuming the data obtained by the light sensor is L=400lx<500lx, the camera exposure is increased by 30%; assuming the data obtained by the noise sensor is N=75dB<80dB, voice interaction is enabled normally. In addition, OpenCV is used to desensitize the camera image; the wearable device heart rate data is stored using a UUID to generate a user ID, which is associated with the real name and set to be viewable only by the administrator. In addition, access permissions are set: the administrator can modify the sensor parameters and view all data, caregivers can only view the data of their assigned area, and visitors can only view the behavior statistics of the public area.

[0081] Next, the sensors were activated to acquire biological behavior data, and the acquired biological behavior data was preprocessed to generate standard behavior data. During data preprocessing, the mean μ=5m and standard deviation σ=0.8m of the infrared sensor position data were calculated, and outliers <5-3×0.8=2.6m or >5+3×0.8=7.4m were removed, with a retention rate of 96%. In addition, data was synchronized based on GPS timestamps (accuracy 0.5ms) to ensure that the time deviation between the camera action data and the wearable device heart rate data was ≤8ms. The heart rate of 40-180 beats / minute was mapped to the 0-1 interval (e.g., 60 beats / minute corresponds to 0.14), and the limb movement amplitude of 0°-180° was mapped to the 0-1 interval (e.g., 90° corresponds to 0.5), and the standard behavior dataset in JSON format was output.

[0082] Next, local features of the standard behavioral data are extracted, specifically including the slope of the movement trajectory (e.g., 0.2 for slow walking in older adults), the rate of change of joint angles (e.g., 0.3° / ms for knee joint changes when going up and down stairs), and the coefficient of variation of heart rate (e.g., 0.05 at rest). A dynamic weighting formula is then used. Calculate the weights and the confidence level of the data source at a given time location. =0.9 (no missing data, 92% match with other data, 3% fluctuation), action data source =0.85 (no missing data, 88% match, 4% fluctuation), physiological data source =0.8 (no missing data, 85% match, 5% fluctuation), calculated as follows =0.9 / (0.9+0.85+0.8)=0.35, =0.85 / 2.55≈0.33, =0.8 / 2.55≈0.32; Combine the CNN-LSTM model for feature fusion, retain the features with confidence ≥0.85, and output a 32-dimensional fusion vector.

[0083] Next, a typical behavior database is accessed, and a neural network classification model is used to identify the behavior category. For example, if an elderly person is found to be stationary in a room for 35 minutes, the category "sedentary" is output with a confidence level of 0.92 and a recognition delay of 400ms. When an elderly person's movement at the stairwell suddenly increases by 50° and then by 3° over the next 10 seconds, it is determined to be a "sudden fall." After a sudden fall is detected, an alert is sent to the caregiver's mobile app at a "high risk" level with a response time of 250ms. At the same time, a buzzer and a red LED light at the stairwell are triggered until the caregiver clicks "handled."

[0084] After the behavior recognition process is complete, a structured report is output to the administrator's PC terminal, generating a 2D heat map of the common activity area and a 3D trajectory animation showing the elderly person's walking path from their room to the dining room. It also connects to the nursing home management platform via a RESTful API.

[0085] When a caregiver reports via touchscreen that "slow walking" is misidentified as "standing", the system marks this data as a training sample. After incremental learning and updating the classification model of the behavior recognition module, the accuracy of "slow walking" recognition improves by 0.8%. When the frequency of "falling" behavior feedback reaches 5 times, the system pushes an optimization suggestion to the administrator to "supplement samples with different fall angles".

[0086] While the present invention has been illustrated and described with reference to certain preferred embodiments, those skilled in the art should understand that the above description is a further detailed explanation of the invention in conjunction with specific embodiments, and should not be construed as limiting the specific implementation of the invention to these descriptions. Various changes in form and detail can be made by those skilled in the art, including several simple deductions or substitutions, without departing from the spirit and scope of the invention.

Claims

1. A behavior perception system, characterized in that, include: A data generation module acquires biological behavior data, preprocesses the biological behavior data to generate standard behavior data, and sends the standard behavior data. A data fusion module receives the standard behavioral data from the data generation module, extracts local features from the standard behavioral data, performs fusion processing on the local features to generate a fusion feature vector, and sends the fusion feature vector. The behavior recognition module stores a typical behavior library. The behavior recognition module receives the fused feature vector from the data fusion module, performs recognition processing on the fused feature vector according to the typical behavior library to output the corresponding behavior category and sends the behavior category; and the behavior recognition module receives the standard behavior data from the data generation module, generates behavior attributes according to the standard behavior data and the behavior category and sends the behavior attributes. The result output module receives the behavior category and the behavior attribute from the behavior recognition module, and generates a recognition report based on the behavior category and the behavior attribute.

2. The behavior perception system as described in claim 1, characterized in that, The data generation module includes a data acquisition unit, a data preprocessing unit, and a data quality monitoring unit; and The biological behavioral data includes location data, movement data, and physiological data; in The data acquisition unit includes a location data acquisition unit for acquiring the location data, a motion data acquisition unit for acquiring the motion data, a physiological data acquisition unit for acquiring the physiological data, and a supplementary data acquisition unit for acquiring supplementary location data and biological movement data. and The data preprocessing unit receives the biological behavior data from the data acquisition unit and preprocesses the biological behavior data to generate the standard behavior data. in The preprocessing includes at least one of the following: removing abnormal data, data synchronization, and data format standardization; and The data quality monitoring unit monitors the signal strength, data packet loss rate, and power supply status of the location data collector, the motion data collector, and the physiological data collector. When the signal strength, data packet loss rate, or power supply status of any collector is abnormal, the unit sends a control command to the data acquisition unit to activate the supplementary data collector. Furthermore, the data quality monitoring unit stores and sends the abnormal information of the collector, including the failure time of the location data collector, the motion data collector, and the physiological data collector, as well as the number of times the supplementary data collector has been activated.

3. The behavior perception system as described in claim 2, characterized in that, When the signal strength, data packet loss rate, or power supply status corresponding to any data acquisition unit is abnormal, a control command is sent to the data acquisition unit to activate the supplementary data acquisition unit, including: When the location data collector malfunctions, the supplementary data collector is activated to obtain supplementary location data and supplementary movement speed data of the organism; and, after activating the supplementary data collector, the following steps are also included: The standard behavior data is generated based on the supplementary location data and the supplementary movement speed data; When the motion data collector malfunctions, the supplementary data collector is activated to acquire the supplementary location data; and after activating the supplementary data collector, the process further includes: The motion amplitude is calculated based on the supplementary location data acquired by the supplementary data collector, the position change is calculated based on the location data, and the standard behavior data is generated based on the motion amplitude and the position change; and... When the physiological data acquisition device malfunctions, the supplementary data acquisition device is activated to obtain the supplementary location data; and after activating the supplementary data acquisition device, the process further includes: The motion amplitude is calculated based on the supplementary location data acquired by the supplementary data collector, and the standard behavior data is inferred based on the motion amplitude and the motion data.

4. The behavior perception system as described in claim 2, characterized in that, The data fusion module extracts local features from the standard behavioral data, including: Extract the slope of the movement trajectory from the location data, the rate of change of joint angles from the motion data, and the coefficient of variation of heart rate from the physiological data; and When the data preprocessing unit generates the standard behavioral data, it also includes: Data confidence levels are generated for the location data, action data, and physiological data based on the completeness, consistency, and stability of the standard behavioral data. The data fusion module performs fusion processing on the local features, including: Based on the data confidence level, initial weights are calculated for the location data, the action data, and the physiological data using a weighted fusion algorithm. Then, the local features are fused according to these initial weights to generate the fused feature vector. The data fusion module calculates the initial weights according to the following formula: in, For the first The weights of the data types range from 0 to 1, and the initial sum of the weights of the location data, the action data, and the physiological data is 1. For the data confidence level of class X; and The data fusion module further includes a data consistency monitoring unit; wherein The data consistency monitoring unit generates a consistency judgment result for the location data, the action data, and the physiological data, and sends control information to the data generation module to reacquire the biological behavior data when the judgment result is that the data is inconsistent.

5. The behavior perception system as described in claim 4, characterized in that, The behavior recognition module uses a neural network classification model to identify and process the fused feature vector and outputs the behavior category; and The behavior recognition module also outputs a category confidence score corresponding to the behavior category; and The behavioral attributes include the time of occurrence, location of occurrence, duration of duration, and frequency of the behavior; and The behavior recognition module also includes an anomaly warning unit; in The anomaly warning unit stores preset abnormal behavior standards, and generates tiered warning information based on the behavior category, the behavior attribute, and the abnormal behavior standards; and The recognition report generated by the result output module includes at least one of the following: a structured behavior report, a two-dimensional heatmap, and a three-dimensional behavior trajectory animation; and The result output module also includes an external protocol port, through which the result output module sends the identification report to the target device.

6. The behavior perception system as described in any one of claims 1-5, characterized in that, It also includes a data security module, which comprises an encryption management unit, a permission setting unit, a behavior recording unit, a storage unit, and a privacy protection unit; wherein The encryption management unit is connected to the data generation module, the data fusion module, the behavior recognition module, and the result output module, respectively, and performs encryption and decryption processing on the data when the data generation module, the data fusion module, the behavior recognition module, and the result output module send and receive data; The permission setting unit sets operation permissions for all data, including permissions that can only be operated by administrators, permissions that can only be operated by administrators and operators, and permissions that can be made public to visitors. The behavior recording unit records all operations corresponding to each piece of data and generates a behavior log, and saves the behavior log in the read-only partition of the storage unit; The privacy protection unit desensitizes the action data in the standard behavior data and anonymizes the standard behavior data corresponding to each organism, generates an identification code for each organism, and stores the identification code and the corresponding standard behavior data in the storage unit.

7. The behavior perception system as described in any one of claims 1-5, characterized in that, It also includes a feedback optimization module, which includes a touch interaction unit and / or a voice interaction unit, as well as an optimization unit; in The touch interaction unit receives externally input text interaction information and / or gesture interaction information; The voice interaction unit receives externally input voice interaction information; The optimization unit extracts feedback information based on the text interaction information, the gesture interaction information, and / or the voice interaction information, and optimizes the neural network classification model of the behavior recognition module based on the feedback information.

8. The behavior perception system as described in claim 7, characterized in that, It also includes a data correction module, which comprises an environmental data acquisition unit and a correction unit; wherein The environmental data acquisition unit acquires temperature data, humidity data, light data, and sound data of the target area. The environmental data acquisition unit includes a temperature and humidity sensor for acquiring the temperature data and humidity data, a light sensor for acquiring the light data, and a noise sensor for acquiring the sound data. The correction unit corrects the position information and / or motion information in the standard behavior data based on the temperature data and the humidity data, and using the following formula: in, For the corrected location information and / or motion information; The raw data acquired by the temperature and humidity sensor; The temperature influence coefficient ranges from 0.005 to 0.01, indicating that the correction range for the data is 0.5% to 1% for every 1°C deviation of the temperature from the standard value; T represents the temperature data in °C. The standard temperature is 25℃. The humidity influence coefficient ranges from 0.003 to 0.008, indicating that the correction margin for the data is 0.3% to 0.8% for every 1% RH deviation of the humidity from the standard value. The humidity data is in %RH. The standard humidity is 50%RH. The correction unit corrects the exposure parameters of the motion data acquisition device in the data generation module based on the illumination data; wherein, when the illumination data is less than or equal to a preset illumination threshold, the exposure of the motion data acquisition device is increased. The correction unit corrects the voice interaction unit of the feedback optimization module based on the sound data; wherein, when the sound data is greater than a preset sound threshold, the unit stops receiving input data from the voice interaction unit.

9. A behavior perception method applicable to the behavior perception system as described in any one of claims 1-8, characterized in that, include: Acquire biological behavior data and preprocess the biological behavior data to generate standard behavior data; Local features are extracted from the standard behavioral data, and the local features are fused to generate a fused feature vector; Based on a typical behavior library, the fused feature vector is identified using a neural network classification model, and the corresponding behavior category is output. Behavioral attributes are generated based on the standard behavior data and the behavior category. An identification report is generated based on the behavior category and the behavior attribute.

10. The behavior perception method as described in claim 9, characterized in that, Also includes: The biological behavior data, the standard behavior data, the fused feature vector, the behavior category, the behavior attribute, and the identification report are all encrypted and operation permissions are set. and / or A behavior log is generated based on all operations corresponding to the biological behavior data, the standard behavior data, the fused feature vector, the behavior category, the behavior attribute, and the identification report. and / or Acquire externally input text interaction information, gesture interaction information, and / or voice interaction information, extract feedback information based on the text interaction information, gesture interaction information, and / or voice interaction information, and optimize the neural network classification model based on the feedback information; and / or Acquire environmental data of the target area, and correct the standard behavior data and the voice interaction information based on the environmental data.

Images