A portable real-time image analysis work system

Abstract

The application discloses a kind of portable real-time image analysis work systems, it is related to computer hardware and image analysis technical field, including: by automatic identification and extraction image frame data in hard disk, establish the unique association of data and storage device to ensure traceable source;Real-time monitoring is carried out to image content based on deep learning algorithm, identifies abnormal behavior, and abnormal image frame is stored by automatic interception, compression;Image frame data and associated information are managed using structured approach, supporting efficient retrieval and backtracking;Through collaborative scheduling module, dynamically allocate computing power, optimize the operating efficiency of each module, while combining user feedback to continuously update the abnormal behavior feature library, improve the identification accuracy.The application has the advantages that: it realizes non-intervention image extraction and accurate traceability, real-time intelligent identification of abnormalities and retention of key image frames, while having efficient data backtracking, flexible system collaboration and iterative optimization capabilities, adapting to the efficient data processing needs of portable scenarios.

Description

Technical Field

[0001] This invention relates to the fields of computer hardware and image analysis technology, and in particular to a portable real-time image analysis system. Background Technology

[0002] Portable real-time image analysis workstations are highly efficient intelligent devices that have emerged in recent years with the development of artificial intelligence, edge computing, and mobile imaging technologies. They are mainly designed for real-time image processing needs in fields such as security patrol, emergency rescue, industrial quality inspection, and media production.

[0003] Currently available portable real-time image analysis workstations generally suffer from several shortcomings. Firstly, they lack automation and adaptability. Most systems require manual intervention for file selection and format parsing after hard drive access, exhibiting weak multi-format compatibility. Data extraction is easily incomplete due to image corruption or encoding incompatibility, and efficient integrity verification mechanisms are lacking. Secondly, traceability capabilities are weak or absent. It's difficult to establish a unique link between image frame data and the original storage hard drive, hindering accurate source tracing and failing to meet the closed-loop evidence chain requirements of scenarios such as judicial evidence collection. Thirdly, anomaly analysis and data processing efficiency are limited. Some systems rely on traditional algorithms, resulting in low accuracy in anomaly recognition, high rates of missed and false positives, and fixed feature libraries that cannot be dynamically optimized. Image frame capture often follows a fixed duration pattern, lacking targeted intelligent compression strategies, leading to high storage resource consumption and inaccurate retention of key information. Fourthly, system collaboration and backtracking efficiency are low. Module interoperability is poor, computing power allocation is rigid, and lag is common. A multi-dimensional indexing system is lacking, data retrieval relies on single conditions, backtracking is time-consuming, and the display of related information is not intuitive, making it difficult to adapt to the high-efficiency operation requirements of portable scenarios. Summary of the Invention

[0004] To improve existing workstations, a portable real-time image analysis system is provided. This workstation combines automation and intelligence, enabling non-interventional image extraction and accurate source tracing, real-time intelligent anomaly identification and retention of key image frames, and also possessing efficient data backtracking, flexible system collaboration and iterative optimization capabilities, adapting to the efficient data processing needs of portable scenarios.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A portable real-time image analysis system includes:

[0007] Data recognition and reading module: used to extract image frame data without manual intervention after hard drive connection, and to obtain valid image frames;

[0008] Hard disk ID association and traceability module: used to establish a unique association between image frame data and the original storage hard disk, and to trace the source of the data;

[0009] Anomaly analysis module: Based on deep learning target detection algorithms, it performs real-time anomaly identification on image frame sequences, including motion trajectories, specific actions, and signals;

[0010] Image capture and compression module: Based on the anomaly trigger signal, it captures and stores anomaly-related image frames using a timestamp anchoring algorithm;

[0011] Data association storage module: Stores image frame data and various associated information in a structured manner, and establishes indexes and backups;

[0012] Data backtracking module: Based on multi-condition parsing and matching, it quickly retrieves and backtracks target image frame data to obtain related information data;

[0013] System Coordination and Scheduling Module: Used to coordinate the workflow of each module, generate and issue instructions, and dynamically allocate the overall computing power of the system;

[0014] Abnormal Behavior Feature Update Module: Updates the feature library based on user feedback data to dynamically optimize the ability to identify abnormal behaviors.

[0015] Preferably, the data recognition and reading module specifically includes:

[0016] Hard drive access sensing unit: Real-time monitoring of the device connection status of the access port, and triggering the device identification process when a hard drive is detected to be connected;

[0017] Multi-format compatible parsing unit: performs a traversal scan of files in the hard disk storage medium, and parses the encoding format, resolution, size, and container format parameters of image files;

[0018] Integrity verification unit: performs integrity verification on the identified image files and removes damaged, incomplete or undecodeable image frame data;

[0019] Valid image frame extraction unit: Based on the parsing results and verification conclusions, it extracts valid image frames that meet the requirements and temporarily stores the extracted image frame data through a data caching mechanism.

[0020] Preferably, the hard disk ID association and tracing module specifically includes:

[0021] Hard disk ID acquisition unit: reads the unique identifier of the connected hard disk through a hardware interface protocol. The identifier is unique characteristic information, including hard disk serial number and physical address.

[0022] Association information generation unit: binds the acquired hard disk IDs to the extracted image frames one by one, and generates a source association dataset containing hard disk IDs, image frame sequence data start timestamps, image frame sequence data end timestamps, and data storage path indexes;

[0023] Related information synchronization unit: synchronizes the generated source tracing related dataset to the data association storage module in real time, and stores the source tracing information and the corresponding image frame data synchronously;

[0024] The tracing unit can reversely query the corresponding hard disk ID by inputting the association information of the target image frame data, or query the image frame data and circulation records stored on the hard disk by inputting the hard disk ID.

[0025] Preferably, the abnormal behavior analysis module specifically includes:

[0026] Detection and Tracking Unit: Real-time detection of target objects in preprocessed image frames is performed using a deep learning target detection algorithm, marking the position coordinates and contour data of the target objects, and continuous tracking is performed based on the appearance features of the target objects using a multi-target tracking algorithm to obtain the motion trajectory data of the target objects;

[0027] Anomaly detection unit for motion trajectory: It presets the normal motion trajectory range, movement speed threshold and motion direction constraints of the target object under different scenarios, compares the real-time acquired motion trajectory data with the preset standards, and determines whether the motion trajectory is abnormal.

[0028] Abnormal Action Recognition Unit: Constructs an abnormal action feature library containing common abnormal action feature parameters, including climbing, running, illegal intrusion, leaving objects behind, and violent conflict. Extracts key feature points of the target limbs through a feature point matching algorithm, calculates similarity with the abnormal action features in the feature library, and determines whether a specific abnormal action exists.

[0029] Abnormal signal triggering unit: When an abnormal motion trajectory or a specific abnormal action is detected, an abnormal triggering signal is immediately generated, including the timestamp of the abnormal occurrence, the abnormal type, and the target object identifier.

[0030] Preferably, the image cropping and compression module specifically includes:

[0031] Extracting time units: Extracting the timestamp of the anomaly based on the anomaly trigger signal, and then tracing back 30 seconds before and after the timestamp to obtain the time interval;

[0032] Image cropping unit: Based on a defined time interval, the corresponding image is accurately cropped from the image frame using a timestamp anchoring algorithm;

[0033] Compression processing unit: It adopts a content-aware intelligent compression algorithm to compress the captured key image frames. It uses a high-fidelity compression strategy for the core area where abnormal behavior occurs and a moderate compression strategy for non-core areas.

[0034] Image association and transmission unit: binds the compressed key image frames with source tracing association information and anomaly analysis results, obtains the complete data packet, and transmits it to the data association storage module for archiving.

[0035] Preferably, the data association storage module specifically includes:

[0036] Database building unit: Through a relational database architecture, a database containing multi-dimensional data fields is built, including: disk ID field, image frame unique identifier field, image frame original information field, anomaly type field, anomaly timestamp field, key image frame storage path field, source tracing association information field, and data storage time field;

[0037] Categorized storage unit: Data is categorized and entered based on database field definitions. Image frame data is stored in binary form in a designated storage area, and associated information is stored in structured data form.

[0038] Index building unit: A multi-dimensional index system is built based on disk ID, exception timestamp, and exception type, and retrieval efficiency is optimized through B+ tree index structure;

[0039] Data security backup unit: It adopts a dual backup mechanism to perform off-site backup and local backup of compressed key image frames and traceability information.

[0040] Preferably, the data backtracking module specifically includes:

[0041] Search condition parsing and matching unit: Converts natural language search conditions into query statements that the database can recognize, and filters and matches structured data in the database based on a preset index system;

[0042] Target image frame retrieval unit: Based on the matched data record, extract the corresponding image frame data storage path and retrieve the corresponding image frame data;

[0043] Related information display unit: When retrieving image frame data, the corresponding source related information and anomaly analysis results are extracted and displayed synchronously with the image frame data in the form of a timeline and tab bar.

[0044] Preferably, the system collaborative scheduling module specifically includes:

[0045] Module status monitoring unit: collects real-time working status data of each functional module, including data reading progress, analysis and computing load, and storage resource usage, and establishes a module status monitoring ledger;

[0046] Command generation and issuance unit: Generates scheduling commands based on the status data of each module: When the data recognition and reading module completes the extraction of image frame sequences, it issues an analysis start command to the abnormal behavior analysis module. When the analysis module detects an anomaly and generates a trigger signal, it issues a collaborative execution command to the image capture and compression module and the data association storage module.

[0047] Computing resource allocation unit: Real-time monitoring of the overall computing load of the system, using a load balancing algorithm to prioritize the allocation of computing resources to the abnormal behavior analysis module and the image capture module, and dynamically adjusting the resource usage of non-core modules when computing power is strained;

[0048] Anomaly Handling Unit: When an anomaly is detected in a module, the fault tolerance mechanism is triggered. The process is restored by retransmitting data, restarting the module, or switching to a backup processing link, and the anomaly log is recorded.

[0049] Preferably, the abnormal behavior feature update module specifically includes:

[0050] Feedback Data Unit: Collects user feedback on anomaly analysis results, including misjudgment annotations, omissions, and requests for new anomaly types;

[0051] Feature library update unit: Based on the collected feedback data, the abnormal behavior feature library is optimized using machine learning iterative algorithms. This includes feature extraction and analysis of misjudged cases, removal of redundant feature parameters that cause misjudgment from the feature library, adjustment of feature matching weights, feature learning for missed abnormal behaviors, extraction of key feature parameters of such abnormal behaviors, addition to the abnormal behavior feature library, and definition of feature standards for such abnormal types according to the requirements of new abnormal types, which are then added to the feature library.

[0052] The identification threshold optimization unit dynamically adjusts the similarity threshold and trajectory deviation threshold parameters for abnormal behavior identification using a gradient descent algorithm, and calculates the optimal threshold range based on feedback data.

[0053] Compared with the prior art, the advantages of the present invention are:

[0054] With outstanding automation and intelligence, it can extract and verify the integrity of image frames without manual intervention after the hard drive is connected, and ensure the accuracy of data traceability through a unique hard drive ID association mechanism. The deep learning-based abnormal behavior analysis module can accurately identify trajectory anomalies and specific dangerous actions in real time. Combined with timestamp-anchored intelligent cropping and compression technology, it can efficiently retain key image frames and optimize storage resources. Multi-dimensional indexed structured storage and fast backtracking function improve data retrieval efficiency. The system collaborative scheduling module realizes efficient linkage and dynamic allocation of computing power among various links, and supports iterative optimization of feature library based on user feedback to continuously improve recognition accuracy. Overall, it takes into account efficient data processing, reliable traceability and flexible adaptability in portable scenarios. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the system proposed in this invention;

[0056] Figure 2 This is a diagram of the data recognition and reading module proposed in this invention;

[0057] Figure 3 This is a diagram of the hard disk ID association and tracing module proposed in this invention;

[0058] Figure 4 This is a diagram of the abnormal behavior analysis module proposed in this invention;

[0059] Figure 5 This is a diagram of the image cropping and compression module proposed in this invention;

[0060] Figure 6 This is a diagram of the data association storage module proposed in this invention;

[0061] Figure 7 This is a diagram of the data backtracking module proposed in this invention;

[0062] Figure 8 This is a diagram of the system collaborative scheduling module proposed in this invention;

[0063] Figure 9 This is a diagram of the abnormal behavior feature update module proposed in this invention. Detailed Implementation

[0064] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0065] See Figure 1 As shown, a portable real-time image analysis system includes:

[0066] Data recognition and reading module: used to extract image frame data without manual intervention after hard drive connection, and to obtain valid image frames;

[0067] Hard disk ID association and traceability module: used to establish a unique association between image frame data and the original storage hard disk, and to trace the source of the data;

[0068] Anomaly analysis module: Based on deep learning target detection algorithms, it performs real-time anomaly identification on image frame sequences, including motion trajectories, specific actions, and signals;

[0069] Image capture and compression module: Based on the anomaly trigger signal, it captures and stores anomaly-related image frames using a timestamp anchoring algorithm;

[0070] Data association storage module: Stores image frame data and various associated information in a structured manner, and establishes indexes and backups;

[0071] Data backtracking module: Based on multi-condition parsing and matching, it quickly retrieves and backtracks target image frame data to obtain related information data;

[0072] System Coordination and Scheduling Module: Used to coordinate the workflow of each module, generate and issue instructions, and dynamically allocate the overall computing power of the system;

[0073] Abnormal Behavior Feature Update Module: Updates the feature library based on user feedback data to dynamically optimize the ability to identify abnormal behaviors.

[0074] See Figure 2 As shown, the data recognition and reading module specifically includes:

[0075] Hard drive access sensing unit: Real-time monitoring of the device connection status of the access port, and triggering the device identification process when a hard drive is detected to be connected;

[0076] Multi-format compatible parsing unit: performs a traversal scan of files in the hard disk storage medium, and parses the encoding format, resolution, size, and container format parameters of image files;

[0077] Integrity verification unit: performs integrity verification on the identified image files and removes damaged, incomplete or undecodeable image frame data;

[0078] Valid image frame extraction unit: Based on the parsing results and verification conclusions, it extracts valid image frames that meet the requirements and temporarily stores the extracted image frame data through a data caching mechanism.

[0079] Specifically, a depth-first search strategy is used to traverse the entire directory structure of the hard drive, automatically skipping system hidden files, temporary cache files and non-image format files; at the same time, the file metadata area is parsed to obtain core parameters such as resolution, size, bitrate and container format;

[0080] A circular buffer is used for temporary data storage, and the buffer capacity dynamically adapts to the image file size. During the extraction process, the valid image frames are segmented and marked according to the image parameter file, and the start timestamp and data offset of each segment are recorded synchronously.

[0081] See Figure 3 As shown, the hard disk ID association and tracing module specifically includes:

[0082] Hard disk ID acquisition unit: reads the unique identifier of the connected hard disk through a hardware interface protocol. The identifier is unique characteristic information, including hard disk serial number and physical address.

[0083] Association information generation unit: binds the acquired hard disk IDs to the extracted image frames one by one, and generates a source association dataset containing hard disk IDs, image frame sequence data start timestamps, image frame sequence data end timestamps, and data storage path indexes;

[0084] Related information synchronization unit: synchronizes the generated source tracing related dataset to the data association storage module in real time, and stores the source tracing information and the corresponding image frame data synchronously;

[0085] The tracing unit can reversely query the corresponding hard disk ID by inputting the association information of the target image frame data, or query the image frame data and circulation records stored on the hard disk by inputting the hard disk ID.

[0086] Specifically, using the collected unique hard disk ID as the core index, the full set of key information of the corresponding image frame is associated to generate a structured traceability and association dataset, which includes: basic hard disk information, core image parameters, time dimension information, and storage location information; a bidirectional mapping mechanism of "hard disk ID - unique image frame identifier ID" is adopted to assign a unique image frame identifier ID to each image frame segment;

[0087] When querying by hard drive ID, the user enters the target hard drive ID, and based on a preset multi-dimensional index, matches all unique identifiers of image frames associated with that hard drive ID in the database. It then retrieves the corresponding image core parameters, time information, and storage path, and outputs a list of all image frames stored on that hard drive and their transfer records. When querying by image information, the user can enter any one or more pieces of information, such as the unique identifier of the image frame. Through a combination of fuzzy matching and precise retrieval, the corresponding image frame data is located, and its associated hard drive ID and basic hard drive information are extracted in reverse association, presenting the original storage medium of the image frame data.

[0088] See Figure 4 As shown, the abnormal behavior analysis module specifically includes:

[0089] Detection and Tracking Unit: Real-time detection of target objects in preprocessed image frames is performed using a deep learning target detection algorithm, marking the position coordinates and contour data of the target objects, and continuous tracking is performed based on the appearance features of the target objects using a multi-target tracking algorithm to obtain the motion trajectory data of the target objects;

[0090] Anomaly detection unit for motion trajectory: It presets the normal motion trajectory range, movement speed threshold and motion direction constraints of the target object under different scenarios, compares the real-time acquired motion trajectory data with the preset standards, and determines whether the motion trajectory is abnormal.

[0091] Abnormal Action Recognition Unit: Constructs an abnormal action feature library containing common abnormal action feature parameters, including climbing, running, illegal intrusion, leaving objects behind, and violent conflict. Extracts key feature points of the target limbs through a feature point matching algorithm, calculates similarity with the abnormal action features in the feature library, and determines whether a specific abnormal action exists.

[0092] Abnormal signal triggering unit: When an abnormal motion trajectory or a specific abnormal action is detected, an abnormal triggering signal is immediately generated, including the timestamp of the abnormal occurrence, the abnormal type, and the target object identifier.

[0093] Specifically, a lightweight deep learning object detection algorithm is adopted to meet real-time processing requirements. Standardized image frames are trained and scanned frame by frame, focusing on detecting personnel, vehicles, and carried items. During the detection process, a sliding window is used to traverse the image frame sequence, extracting image features within each window and comparing them with a preset target feature library to generate the bounding box coordinates, confidence score, and category label of the target object. Only detection results with a confidence score higher than 95% are retained. In the tracking stage, appearance features and motion features are fused to extract the color histogram, texture features, and contour features of the target object. The Kalman filter algorithm is used to predict the target's motion trajectory. When the target is briefly occluded, the position is continuously predicted based on historical motion data and surrounding environmental features. After the occlusion is removed, the target is quickly re-matched. At the same time, the target bounding boxes of adjacent frames are calculated through intersection-union ratio (IoU) to update the target tracking ID and generate complete trajectory data containing the target ID, position coordinates of each frame, movement speed, and direction of movement.

[0094] Based on different application scenarios, a differentiated normal movement rule base is preset. For closed and controlled areas, polygonal geofences are defined as the boundaries for prohibited entry and exit. For passage-type scenarios, one-way or two-way traffic direction constraints are set. For personnel activity areas, a normal movement speed range of 0.5-5 m / s is set. The real-time generated trajectory data is compared with the rule base frame by frame: if the target trajectory point exceeds the geofence boundary for 3 consecutive frames, it is judged as "entry into the restricted area" or "illegal departure"; if the movement speed is below 0.5 m / s or above 5 m / s for 5 consecutive frames, it is judged as "abnormal loitering" or "rapid escape"; if the movement direction is at an angle of more than 90 degrees to the passage constraint for 2 consecutive frames, it is judged as "reverse passage".

[0095] An abnormal action feature library covering core scenarios was constructed, including 12 common abnormal actions such as climbing, vaulting, illegal entry, leaving objects behind, physical conflict, and carrying contraband. Each action category was labeled with 30-50 key feature points. A key point detection algorithm was used to extract 24 key limb nodes of the target object, including the head, shoulders, elbows, wrists, hips, knees, and ankles. The relative distance, angle changes, and movement trajectories between nodes were calculated to generate action feature vectors. The similarity of this vector was compared with the standard vectors of various abnormal actions in the feature library. When the similarity exceeded 85% and lasted for more than 3 frames, the corresponding abnormal action was determined to exist.

[0096] The formula for calculating the similarity of abnormal action features is:

[0097]

[0098] in, The feature similarity between the action to be detected and the standard abnormal action. Let be the feature vector of the action to be detected. Let i be the eigenvalue of the i-th dimension of the vector. This is the feature vector of a standard abnormal action. Let i be the standard eigenvalue of the i-th dimension of the vector. is the dimension of the feature vector.

[0099] See Figure 5 As shown, the image cropping and compression module specifically includes:

[0100] Extracting time units: Extracting the timestamp of the anomaly based on the anomaly trigger signal, and then tracing back 30 seconds before and after the timestamp to obtain the time interval;

[0101] Image cropping unit: Based on a defined time interval, the corresponding image is accurately cropped from the image frame using a timestamp anchoring algorithm;

[0102] Compression processing unit: It adopts a content-aware intelligent compression algorithm to compress the captured key image frames. It uses a high-fidelity compression strategy for the core area where abnormal behavior occurs and a moderate compression strategy for non-core areas.

[0103] Image association and transmission unit: binds the compressed key image frames with source tracing association information and anomaly analysis results, obtains the complete data packet, and transmits it to the data association storage module for archiving.

[0104] Specifically, a timestamp-anchored capture method is adopted. Based on a defined time interval, the corresponding frame sequence is quickly located through the timestamp index of the image frame. During the capture process, the integrity of the frame data is monitored in real time, and the frame header identifier, data length and check code are verified frame by frame. If a frame is found to be incomplete, the image information of the adjacent normal frame is automatically called to complete and repair it, and the repair position is marked for subsequent verification. After the capture is completed, image frame metadata is generated, which includes the unique identifier of the image frame, the original image frame association ID, the capture time range, the encoding format and other information, and is bound to the captured image frame.

[0105] A content-aware hierarchical compression strategy is adopted. First, an abnormal core region and a non-core region in the image frame are identified and extracted through an image segmentation algorithm. For the core region, a lossless compression algorithm with a low compression ratio is used to preserve the pixel-level details of the image. For the non-core region, an adaptive lossy compression algorithm with a medium to high compression ratio is used. At the same time, edge sharpening processing is used to avoid blurring and bokeh after compression of the non-core region.

[0106] See Figure 6 As shown, the data association storage module specifically includes:

[0107] Database building unit: Through a relational database architecture, a database containing multi-dimensional data fields is built, including: disk ID field, image frame unique identifier field, image frame original information field, anomaly type field, anomaly timestamp field, key image frame storage path field, source tracing association information field, and data storage time field;

[0108] Categorized storage unit: Data is categorized and entered based on database field definitions. Image frame data is stored in binary form in a designated storage area, and associated information is stored in structured data form.

[0109] Index building unit: A multi-dimensional index system is built based on disk ID, exception timestamp, and exception type, and retrieval efficiency is optimized through B+ tree index structure;

[0110] Data security backup unit: It adopts a dual backup mechanism to perform off-site backup and local backup of compressed key image frames and traceability information.

[0111] Specifically, a multi-level relational database architecture is built, dividing the database into three logical storage areas: core data area, index area, and backup area. The core data area is designed to cover key information throughout the entire process, including: disk ID, unique image frame identifier ID, original image frame parameter set, anomaly information set, storage path index, source tracing related data, and security identifier. The index area is based on high-frequency retrieval scenarios, establishing three types of composite indexes: "disk ID + anomaly timestamp", "unique image frame identifier ID", and "anomaly type + region coordinates". A B+ tree index structure is used to optimize the efficiency of range queries and precise matching.

[0112] After receiving data transmitted from each module, the system uses a binary streaming storage method for raw image frames, preprocessed image frames, and compressed key image frames. Files are named according to the naming rule of "disk ID - image frame unique identifier ID - segment sequence number". Large files are automatically divided into 2GB / segment fragments for storage, and a fragment check code and associated fragment index are added to the end of each segment. For structured data, it is formatted and entered according to the field definition. Before entering the data, invalid data is removed by field format validation and data logic validation. The data writing progress is recorded in real time during the storage process. An integrity check is triggered every time 50% of the data storage is completed, and the data size and check code are compared with the source data.

[0113] See Figure 7 As shown, the data backtracking module specifically includes:

[0114] Search condition parsing and matching unit: Converts natural language search conditions into query statements that the database can recognize, and filters and matches structured data in the database based on a preset index system;

[0115] Target image frame retrieval unit: Based on the matched data record, extract the corresponding image frame data storage path and retrieve the corresponding image frame data;

[0116] Related information display unit: When retrieving image frame data, the corresponding source related information and anomaly analysis results are extracted and displayed synchronously with the image frame data in the form of a timeline and tab bar.

[0117] Specifically, after receiving the user's input search criteria, the system converts the natural language description into a structured query instruction that the database can recognize. During the parsing process, the system prioritizes high-frequency index fields, first extracting precise matching fields such as hard disk ID and unique image frame ID, then processing range matching fields such as time range and anomaly type, and finally parsing fuzzy matching fields such as target object features. Based on the composite index of "hard disk ID + anomaly timestamp" and "anomaly type + region coordinates" preset by the data association storage module, the system first quickly narrows down the data range through precise fields, and then performs fuzzy feature matching within the filtered dataset. For example, when matching target object features, the system extracts key features from the description and compares them with the target feature summary associated with the image, uses semantic similarity analysis to exclude irrelevant data, and provides real-time feedback on the matching progress.

[0118] Based on the matched data records, the corresponding storage path index is extracted. When retrieving images from large file fragment storage, the fragments are automatically assembled in order of "fragment number" and the checksum of each fragment is verified synchronously. If the main storage data retrieval fails, the system automatically switches to the local backup or off-site backup address for retrieval, and a prompt is displayed to inform the user of the data source. During the retrieval process, a combination of streaming transmission and local caching is used. The first 30 seconds of image frame sequence data are cached, and the system is adapted to mainstream playback formats, supporting functions such as speed playback and frame step viewing.

[0119] See Figure 8 As shown, the system collaborative scheduling module specifically includes:

[0120] Module status monitoring unit: collects real-time working status data of each functional module, including data reading progress, analysis and computing load, and storage resource usage, and establishes a module status monitoring ledger;

[0121] Command generation and issuance unit: Generates scheduling commands based on the status data of each module: When the data recognition and reading module completes the extraction of image frame sequences, it issues an analysis start command to the abnormal behavior analysis module. When the analysis module detects an anomaly and generates a trigger signal, it issues a collaborative execution command to the image capture and compression module and the data association storage module.

[0122] Computing resource allocation unit: Real-time monitoring of the overall computing load of the system, using a load balancing algorithm to prioritize the allocation of computing resources to the abnormal behavior analysis module and the image capture module, and dynamically adjusting the resource usage of non-core modules when computing power is strained;

[0123] Anomaly Handling Unit: When an anomaly is detected in a module, the fault tolerance mechanism is triggered. The process is restored by retransmitting data, restarting the module, or switching to a backup processing link, and the anomaly log is recorded.

[0124] Specifically, a load-aware resource scheduling framework is built to summarize the overall computing load of the system in real time and dynamically allocate resources based on the priority of module business. The abnormal behavior analysis module and the image capture and compression module are listed as core business modules, and their resource supply is prioritized, with a default allocation of 60% CPU computing power, 70% GPU computing resources, and 50% memory capacity. The data association storage and retrieval data backtracking modules are listed as secondary priority, and allocated 30% of hardware resources. The abnormal behavior feature update and backup modules are listed as ordinary priority, occupying the remaining 10% of resources. When the core modules experience high load, the load balancing mechanism is automatically activated: the resource occupation of ordinary priority modules is temporarily reduced, and the released resources are supplemented to the core modules. At the same time, a task sharding strategy is adopted to split the abnormal analysis task into multiple sub-tasks and distribute them to idle computing nodes for parallel processing. When the load returns to normal, the resources are gradually returned to the original allocation ratio.

[0125] See Figure 9 As shown, the abnormal behavior feature update module specifically includes:

[0126] Feedback Data Unit: Collects user feedback on anomaly analysis results, including misjudgment annotations, omissions, and requests for new anomaly types;

[0127] Feature library update unit: Based on the collected feedback data, the abnormal behavior feature library is optimized using machine learning iterative algorithms. This includes feature extraction and analysis of misjudged cases, removal of redundant feature parameters that cause misjudgment from the feature library, adjustment of feature matching weights, feature learning for missed abnormal behaviors, extraction of key feature parameters of such abnormal behaviors, addition to the abnormal behavior feature library, and definition of feature standards for such abnormal types according to the requirements of new abnormal types, which are then added to the feature library.

[0128] The identification threshold optimization unit dynamically adjusts the similarity threshold and trajectory deviation threshold parameters for abnormal behavior identification using the gradient descent algorithm, and calculates the optimal threshold range based on feedback data.

[0129] Specifically, based on the filtered feedback data, an iterative update process for the feature library is initiated. For misjudged cases, feature parameters of normal behavior are extracted and compared with similar features in the abnormal behavior feature library to identify redundant features that led to misjudgments. Redundant parameters are then eliminated through feature importance assessment, and the matching weights of relevant features are adjusted. For missed cases, feature point extraction technology is used to capture the core features of abnormal behavior. Combined with behavioral feature patterns in similar scenarios, 30-50 key feature points are added to the feature library. For new abnormal type requirements, technical and business experts are organized to define the feature standards for the behavior, construct complete feature vectors, and classify and store them in the feature library. At the same time, a feature library version management mechanism is established, generating a version number for each update, recording details of added, modified, and deleted features, and supporting historical version backtracking.

[0130] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0131] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A portable real-time image analysis system, characterized in that, include: Data recognition and reading module: used to extract image frame data without manual intervention after hard drive connection, and to obtain valid image frames; Hard disk ID association and traceability module: used to establish a unique association between image frame data and the original storage hard disk, and to trace the source of the data; Anomaly analysis module: Based on deep learning target detection algorithms, it performs real-time anomaly identification on image frame sequences, including motion trajectories, specific actions, and signals; Image capture and compression module: Based on the anomaly trigger signal, it captures and stores anomaly-related image frames using a timestamp anchoring algorithm; Data association storage module: Stores image frame data and various associated information in a structured manner, and establishes indexes and backups; Data backtracking module: Based on multi-condition parsing and matching, it quickly retrieves and backtracks target image frame data to obtain related information data; System Coordination and Scheduling Module: Used to coordinate the workflow of each module, generate and issue instructions, and dynamically allocate the overall computing power of the system; Abnormal behavior feature update module: Updates the feature library based on user feedback data to dynamically optimize the ability to identify abnormal behavior; The data association storage module specifically includes: Database building unit: Through a relational database architecture, a database containing multi-dimensional data fields is built, including: disk ID field, image frame unique identifier field, image frame original information field, anomaly type field, anomaly timestamp field, key image frame storage path field, source tracing association information field, and data storage time field; Categorized storage unit: Data is categorized and entered based on database field definitions. Image frame data is stored in binary form in a designated storage area, and associated information is stored in structured data form. Index building unit: A multi-dimensional index system is built based on disk ID, exception timestamp, and exception type, and retrieval efficiency is optimized through B+ tree index structure; Data security backup unit: It adopts a dual backup mechanism to perform off-site backup and local backup of compressed key image frames and traceability information.

2. The portable real-time image analysis system according to claim 1, characterized in that, The data recognition and reading module specifically includes: Hard drive access sensing unit: Real-time monitoring of the device connection status of the access port, and triggering the device identification process when a hard drive is detected to be connected; Multi-format compatible parsing unit: performs a traversal scan of files in the hard disk storage medium, and parses the encoding format, resolution, size, and container format parameters of image files; Integrity verification unit: performs integrity verification on the identified image files and removes damaged, incomplete or undecodeable image frame data; Valid image frame extraction unit: Based on the parsing results and verification conclusions, it extracts valid image frames that meet the requirements and temporarily stores the extracted image frame data through a data caching mechanism.

3. The portable real-time image analysis system according to claim 1, characterized in that, The hard disk ID association and tracing module specifically includes: Hard disk ID acquisition unit: reads the unique identifier of the connected hard disk through a hardware interface protocol. The identifier is unique characteristic information, including hard disk serial number and physical address. Association information generation unit: binds the acquired hard disk IDs to the extracted image frames one by one, and generates a source association dataset containing hard disk IDs, image frame sequence data start timestamps, image frame sequence data end timestamps, and data storage path indexes; Related information synchronization unit: synchronizes the generated source tracing related dataset to the data association storage module in real time, and stores the source tracing information and the corresponding image frame data synchronously; The tracing unit can reversely query the corresponding hard disk ID by inputting the association information of the target image frame data, or query the image frame data and circulation records stored on the hard disk by inputting the hard disk ID.

4. The portable real-time image analysis system according to claim 1, characterized in that, The abnormal behavior analysis module specifically includes: Detection and Tracking Unit: Real-time detection of target objects in preprocessed image frames is performed using a deep learning target detection algorithm, marking the position coordinates and contour data of the target objects, and continuous tracking is performed based on the appearance features of the target objects using a multi-target tracking algorithm to obtain the motion trajectory data of the target objects; Anomaly detection unit for motion trajectory: It presets the normal motion trajectory range, movement speed threshold and motion direction constraints of the target object under different scenarios, compares the real-time acquired motion trajectory data with the preset standards, and determines whether the motion trajectory is abnormal. Abnormal Action Recognition Unit: Constructs an abnormal action feature library containing common abnormal action feature parameters, including climbing, running, illegal intrusion, leaving objects behind, and violent conflict. Extracts key feature points of the target limbs through a feature point matching algorithm, calculates similarity with the abnormal action features in the feature library, and determines whether a specific abnormal action exists. Abnormal signal triggering unit: When an abnormal motion trajectory or a specific abnormal action is detected, an abnormal triggering signal is immediately generated, including the timestamp of the abnormal occurrence, the abnormal type, and the target object identifier.

5. A portable real-time image analysis system according to claim 1, characterized in that, The image cropping and compression module specifically includes: Extracting time units: Extracting the timestamp of the anomaly based on the anomaly trigger signal, and then tracing back 30 seconds before and after the timestamp to obtain the time interval; Image cropping unit: Based on a defined time interval, the corresponding image is accurately cropped from the image frame using a timestamp anchoring algorithm; Compression processing unit: It adopts a content-aware intelligent compression algorithm to compress the captured key image frames. It uses a high-fidelity compression strategy for the core area where abnormal behavior occurs and a moderate compression strategy for non-core areas. Image association and transmission unit: binds the compressed key image frames with source tracing association information and anomaly analysis results, obtains the complete data packet, and transmits it to the data association storage module for archiving.

6. The portable real-time image analysis system according to claim 1, characterized in that, The data backtracking module specifically includes: Search condition parsing and matching unit: Converts natural language search conditions into query statements that the database can recognize, and filters and matches structured data in the database based on a preset index system; Target image frame retrieval unit: Based on the matched data record, extract the corresponding image frame data storage path and retrieve the corresponding image frame data; Related information display unit: When retrieving image frame data, the corresponding source related information and anomaly analysis results are extracted and displayed synchronously with the image frame data in the form of a timeline and tab bar.

7. A portable real-time image analysis system according to claim 1, characterized in that, The system collaborative scheduling module specifically includes: Module status monitoring unit: collects real-time working status data of each functional module, including data reading progress, analysis and computing load, and storage resource usage, and establishes a module status monitoring ledger; Command generation and issuance unit: Generates scheduling commands based on the status data of each module: When the data recognition and reading module completes the extraction of image frame sequences, it issues an analysis start command to the abnormal behavior analysis module. When the analysis module detects an anomaly and generates a trigger signal, it issues a collaborative execution command to the image capture and compression module and the data association storage module. Computing resource allocation unit: Real-time monitoring of the overall computing load of the system, using a load balancing algorithm to prioritize the allocation of computing resources to the abnormal behavior analysis module and the image capture module, and dynamically adjusting the resource usage of non-core modules when computing power is strained; Anomaly Handling Unit: When an anomaly is detected in a module, the fault tolerance mechanism is triggered. The process is restored by retransmitting data, restarting the module, or switching to a backup processing link, and the anomaly log is recorded.

8. A portable real-time image analysis system according to claim 1, characterized in that, The abnormal behavior feature update module specifically includes: Feedback Data Unit: Collects user feedback on anomaly analysis results, including misjudgment annotations, omissions, and requests for new anomaly types; Feature library update unit: Based on the collected feedback data, the abnormal behavior feature library is optimized using machine learning iterative algorithms. This includes feature extraction and analysis of misjudged cases, removal of redundant feature parameters that cause misjudgment from the feature library, adjustment of feature matching weights, feature learning for missed abnormal behaviors, extraction of key feature parameters of such abnormal behaviors, addition to the abnormal behavior feature library, and definition of feature standards for such abnormal types according to the requirements of new abnormal types, which are then added to the feature library. The identification threshold optimization unit dynamically adjusts the similarity threshold and trajectory deviation threshold parameters for abnormal behavior identification using the gradient descent algorithm, and calculates the optimal threshold range based on feedback data.

Images