An AI visual target detection and behavior recognition intelligent analysis system

Abstract

This invention relates to the field of artificial intelligence visual analysis technology, and discloses an AI visual target detection and behavior recognition intelligent analysis system, including a multi-source visual data acquisition and protocol adaptation module, a side-side data standardization preprocessing and feature enhancement module, a two-stage cascaded AI inference core processing module, an intelligent event analysis and linkage handling module, and a cloud-side data management and model iteration optimization module. This invention adopts a three-level architecture of edge-cloud collaboration, balancing the real-time performance of the system with the convenience of model iteration. By pre-filtering invalid video frames and non-interested regions, it reduces system bandwidth usage and computing power consumption, improving the efficiency of AI inference. Simultaneously, the two-stage cascaded inference architecture balances the real-time performance of target detection with the accuracy of behavior recognition. A multi-dimensional matching mechanism enhances the stability of multi-target tracking, effectively reducing false detections and missed detections during the detection and recognition process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence visual analysis technology, and in particular to an AI visual target detection and behavior recognition intelligent analysis system. Background Technology

[0002] With the rapid development of artificial intelligence technology, AI visual analysis technology has been widely applied in many fields such as smart security, park management, traffic control, and industrial production supervision. Target detection and behavior recognition, as core technologies of AI visual analysis, are key supports for realizing intelligent scene supervision and event handling; Most AI visual intelligent analysis systems currently on the market are deployed and operated using a single architecture. In practical applications, it is difficult to simultaneously ensure the real-time performance, recognition accuracy, and ease of deployment and maintenance of the system. They cannot adapt to the engineering implementation needs of large-scale, multi-scenario applications. Furthermore, they have many shortcomings in data processing, model inference, and iterative optimization, making it difficult to meet the intelligent supervision needs in complex scenarios. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based intelligent analysis system for visual target detection and behavior recognition, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: An AI-powered visual target detection and behavior recognition intelligent analysis system includes a multi-source visual data acquisition and protocol adaptation module, a side-side data standardization preprocessing and feature enhancement module, a two-stage cascaded AI inference core processing module, an intelligent event analysis and coordinated response module, and a cloud-side data management and model iteration optimization module. The output of the multi-source visual data acquisition and protocol adaptation module is connected to the input of the side-side data standardization preprocessing and feature enhancement module; the output of the side-side data standardization preprocessing and feature enhancement module is connected to the input of the two-stage cascaded AI inference core processing module; the output of the two-stage cascaded AI inference core processing module is connected to the input of the intelligent event analysis and coordinated response module; the output of the intelligent event analysis and coordinated response module is connected to the input of the cloud-side data management and model iteration optimization module; and the output of the cloud-side data management and model iteration optimization module is connected to the input of the two-stage cascaded AI inference core processing module.

[0005] As a further improvement to this technical solution: the multi-source visual data acquisition and protocol adaptation module includes multiple types of visual acquisition terminal devices, a multi-protocol compatible access unit, an end-side dynamic frame initial screening unit, and an encrypted transmission and link keep-alive unit; the output end of the multiple types of visual acquisition terminal devices is connected to the input end of the multi-protocol compatible access unit; the output end of the multi-protocol compatible access unit is connected to the input end of the end-side dynamic frame initial screening unit; the output end of the end-side dynamic frame initial screening unit is connected to the input end of the encrypted transmission and link keep-alive unit; the end-side dynamic frame initial screening unit performs grayscale processing on continuously input video frames; to obtain the first... A frame grayscale image is denoted as ;No. A frame grayscale image is denoted as ; Calculate the absolute frame difference between two adjacent frames. ; Calculate the global motion saliency value of the frame difference image. ; Construct adaptive threshold ; Complete frame validity determination; The formula for calculating the adaptive threshold T_t is: , This is a dynamic weighting coefficient; its value ranges from 0.6 to 0.8. These are static weighting coefficients; their values ​​range from 0.2 to 0.4. The preset base threshold; This represents the mean global motion saliency of the previous frame.

[0006] As a further improvement to this technical solution: the side-side data standardization preprocessing and feature enhancement module includes a hardware-accelerated decoding and format standardization unit, a Region of Interest (ROI) filtering unit, and an adaptive image quality enhancement unit; the output of the hardware-accelerated decoding and format standardization unit is connected to the input of the ROI filtering unit; the output of the ROI filtering unit is connected to the input of the adaptive image quality enhancement unit; the hardware-accelerated decoding and format standardization unit completes hardware decoding of the video stream and image format unification processing; the ROI filtering unit completes pixel-level filtering of non-interested regions; and the adaptive image quality enhancement unit completes quality optimization and feature enhancement of the input image.

[0007] As a further improvement to this technical solution: the dual-stage cascaded AI inference core processing module includes a first-stage real-time target detection and multi-target tracking unit, a second-stage spatiotemporal behavior recognition inference unit, and an inference result hierarchical filtering unit; the output of the first-stage real-time target detection and multi-target tracking unit is connected to the input of the second-stage spatiotemporal behavior recognition inference unit; the output of the second-stage spatiotemporal behavior recognition inference unit is connected to the input of the inference result hierarchical filtering unit; the first-stage real-time target detection and multi-target tracking unit completes multi-target detection and cross-frame trajectory association; the trajectory association adopts a matching cost calculation method that fuses Mahalanobis distance and cosine similarity; the formula for calculating the fused matching cost is: , The cost of final fusion matching; This is the weighting coefficient, with a value ranging from 0.2 to 0.5; The Mahalanobis distance; The first stage is the cosine distance; the second stage spatiotemporal dimension behavior recognition reasoning unit completes the temporal feature extraction and category recognition of the target behavior; the reasoning result hierarchical filtering unit completes the deduplication and filtering of reasoning results.

[0008] As a further improvement to this technical solution: the intelligent event analysis and coordinated handling module includes an event classification and rule matching unit, an event tracing and evidence retention unit, a multi-system coordinated handling unit, and an event archiving and statistical analysis unit; the output of the event classification and rule matching unit is connected to the input of the event tracing and evidence retention unit; the output of the event tracing and evidence retention unit is connected to the input of the multi-system coordinated handling unit; the output of the multi-system coordinated handling unit is connected to the input of the event archiving and statistical analysis unit; the event classification and rule matching unit completes the risk level classification and handling process matching of the event; the event tracing and evidence retention unit completes the storage of the corresponding video and key information of the event; the multi-system coordinated handling unit completes the coordinated handling and information push of the event; and the event archiving and statistical analysis unit completes the archiving and statistics of event data.

[0009] As a further improvement to this technical solution: the cloud-side data management and model iteration optimization module includes a scenario-based dataset construction unit, a model incremental training and lightweight optimization unit, a model canary release and OTA upgrade unit, and a global system operation and maintenance and status monitoring unit; the output of the scenario-based dataset construction unit is connected to the input of the model incremental training and lightweight optimization unit; the output of the model incremental training and lightweight optimization unit is connected to the input of the model canary release and OTA upgrade unit; the output of the model canary release and OTA upgrade unit is connected to the input of the global system operation and maintenance and status monitoring unit; the scenario-based dataset construction unit completes the classification storage and label injection of samples into the library; the model incremental training and lightweight optimization unit completes the parameter fine-tuning and structural optimization of the model; the model canary release and OTA upgrade unit completes the version management and push update of the model; and the global system operation and maintenance and status monitoring unit completes the monitoring of the operating status of all system devices and nodes.

[0010] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention adopts a three-level architecture of edge-cloud collaboration, which balances the real-time performance of system operation with the convenience of model iteration. By filtering invalid video frames and non-interested areas in advance, it reduces the system's bandwidth usage and computing power consumption, and improves the operating efficiency of AI inference. At the same time, it adopts a two-stage cascaded inference architecture, which balances the real-time performance of target detection with the accuracy of behavior recognition. Through a multi-dimensional matching mechanism, it improves the stability of multi-target tracking, effectively reducing false detections and missed detections in the detection and recognition process. Through adaptive image enhancement processing, it improves the system's adaptability to different complex environments and ensures the detection and recognition effect in extreme scenarios.

[0011] 2. This invention constructs a complete business closed loop from data collection, analysis and reasoning to event handling and model iteration, which can adapt to the application needs of multiple industries and scenarios. By constructing scenario-based datasets and incrementally training models, it reduces the computing power consumption of model iteration and improves the adaptability of models to specific business scenarios. It realizes batch updates of models through remote upgrades, which significantly reduces the system's operation and maintenance costs. At the same time, through a centralized operation and maintenance management platform, it realizes unified monitoring and management of large-scale deployment nodes, which facilitates the large-scale engineering application of the system.

[0012] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the method structure of an AI visual target detection and behavior recognition intelligent analysis system. Detailed Implementation

[0014] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention. The invention is described more specifically in the following paragraphs by way of example with reference to the accompanying drawings. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0015] Please see Figure 1 In this embodiment of the invention, an AI visual target detection and behavior recognition intelligent analysis system includes a multi-source visual data acquisition and protocol adaptation module, a side-side data standardization preprocessing and feature enhancement module, a two-stage cascaded AI inference core processing module, an intelligent event analysis and coordinated response module, and a cloud-side data management and model iteration optimization module. The output of the multi-source visual data acquisition and protocol adaptation module is connected to the input of the side-side data standardization preprocessing and feature enhancement module; the output of the side-side data standardization preprocessing and feature enhancement module is connected to the input of the two-stage cascaded AI inference core processing module; the output of the two-stage cascaded AI inference core processing module is connected to the input of the intelligent event analysis and coordinated response module; the output of the intelligent event analysis and coordinated response module is connected to the input of the cloud-side data management and model iteration optimization module; and the output of the cloud-side data management and model iteration optimization module is connected to the input of the two-stage cascaded AI inference core processing module. Specifically, the multi-source visual data acquisition and protocol adaptation module completes the acquisition, protocol adaptation, and preliminary screening of on-site visual data, providing effective video stream data for subsequent processing. Side data standardization and preprocessing and feature enhancement module: completes video stream decoding, image format standardization and feature enhancement, providing image data that meets the input requirements for AI inference; The dual-stage cascaded AI inference core processing module completes the core AI inference for target detection, multi-target tracking and behavior recognition, and outputs the recognition results of targets and behaviors. It is the carrier for realizing the core functions of the system. Intelligent event analysis and coordinated response module: completes event matching, hierarchical handling and data archiving of the identification results, realizing a complete business closed loop; Cloud-based data management and model iteration optimization module: Completes dataset construction, model optimization and iteration, and full system operation and maintenance management to achieve continuous system optimization and stable operation; The connection relationship defined in this clause is as follows: each module is connected sequentially according to the data flow direction; the output end of the cloud-side data management and model iteration optimization module is connected to the dual-stage cascaded AI inference core processing module to realize the distribution and iteration of the optimized model, forming a complete edge-cloud collaborative closed loop.

[0016] The multi-source visual data acquisition and protocol adaptation module includes multiple types of visual acquisition terminal devices, a multi-protocol compatible access unit, an end-side dynamic frame initial screening unit, and an encrypted transmission and link keep-alive unit. The output of the multiple types of visual acquisition terminal devices is connected to the input of the multi-protocol compatible access unit; the output of the multi-protocol compatible access unit is connected to the input of the end-side dynamic frame initial screening unit; the output of the end-side dynamic frame initial screening unit is connected to the input of the encrypted transmission and link keep-alive unit; the end-side dynamic frame initial screening unit performs grayscale processing on continuously input video frames to obtain the first... A frame grayscale image is denoted as ;No. A frame grayscale image is denoted as ; Calculate the absolute frame difference between two adjacent frames. ; Calculate the global motion saliency value of the frame difference image. ; Construct adaptive threshold ; Complete frame validity determination; The formula for calculating the adaptive threshold T_t is: , This is a dynamic weighting coefficient; its value ranges from 0.6 to 0.8. These are static weighting coefficients; their values ​​range from 0.2 to 0.4. The preset base threshold; This is the mean of global motion saliency in the previous frame; Specifically, multiple types of visual acquisition terminal devices: complete the real-time acquisition of on-site visual images, and are compatible with multiple types of acquisition terminals such as IPC network cameras, high-speed dome cameras, mobile law enforcement recorders, and drone aerial photography equipment; Multi-protocol compatible access unit: Completes protocol adaptation for different acquisition terminals and stable video stream extraction, supporting mainstream national standards and industry protocols such as RTSP, RTMP, ONVIF, and GB28181; Encrypted transmission and link keep-alive unit: Completes encrypted transmission of video stream data and ensures continuous stability of the communication link, guaranteeing the security and continuity of data transmission; The edge-side dynamic frame initial screening unit filters out static, unchanging, and invalid video frames without effective targets, reducing the computational power consumption and bandwidth usage of subsequent stages. Its implementation logic is motion saliency detection based on adaptive frame difference threshold. It judges whether there are effective dynamic targets in the picture by the pixel changes of consecutive frames, and completes the effective screening of frames. The complete processing flow of this unit is as follows: First, the continuously input video frames are converted to grayscale to obtain the... A frame grayscale image is denoted as , No. A frame grayscale image is denoted as ; Calculate the absolute frame difference between two adjacent frames. ; Calculate the global motion saliency value of the frame difference image. ; Constructed through an adaptive threshold formula Finally, the frame validity determination is completed; when Less than When a frame is deemed invalid, it is filtered out; when Greater than or equal to If a frame is deemed valid, it proceeds to the next processing stage. The standard format of the core adaptive threshold calculation formula is as follows: The formula is labeled and explained. It is used to dynamically adjust the threshold for frame validity judgment, adapting to changes in lighting and environmental disturbances in different scenarios, and avoiding false filtering or missed filtering caused by a fixed threshold. In the formula... For the first The adaptive judgment threshold for a frame is the output value of this formula and is used to determine the validity of subsequent frames. This is a dynamic weighting coefficient, ranging from 0.6 to 0.8, used to balance the influence of the motion significance of the preceding frame on the current threshold. This is a static weighting coefficient, ranging from 0.2 to 0.4, used to balance the influence of the base threshold on the current threshold. and The sum of these values ​​is 1, ensuring the stability of the threshold calculation; The preset basic threshold is a fixed value that is set in advance according to the scene environment to avoid the threshold from failing under extreme conditions. This is the mean global motion saliency of the previous frame, used to characterize the degree of dynamic change in the previous frame.

[0017] The edge data normalization preprocessing and feature enhancement module includes a hardware-accelerated decoding and format normalization unit, a Region of Interest (ROI) filtering unit, and an adaptive image quality enhancement unit. The output of the hardware-accelerated decoding and format normalization unit is connected to the input of the ROI filtering unit; the output of the ROI filtering unit is connected to the input of the adaptive image quality enhancement unit. The hardware-accelerated decoding and format normalization unit performs hardware decoding of the video stream and image format unification processing; the ROI filtering unit performs pixel-level filtering of non-interested regions; and the adaptive image quality enhancement unit performs quality optimization and feature enhancement of the input image. Specifically, the hardware-accelerated decoding and format standardization unit: based on the edge node NPU and GPU hardware acceleration unit, it completes the hardware decoding of H.264 and H.265 format video streams, and at the same time scales the decoded images to the standard size of the model input, and completes standardization processing such as pixel normalization and distortion correction, so as to provide image data with a unified format for subsequent stages; ROI Region Filtering Unit: Based on the spatial mask mapping of the scene prior, it removes invalid background pixels in non-interested areas of the image, reduces the computational load of subsequent AI inference, and improves inference speed. Its implementation logic is to preset the interest area according to the business scenario, construct a spatial mask matrix with the same size as the input image, perform pixel-by-pixel mask filtering on the input image, and retain only the image data of the interest area for subsequent processing. Adaptive Image Quality Enhancement Unit: Optimizes image quality for images in low light, backlight, rain, fog, and slight motion blur, improving the detection and recognition accuracy of low-quality images.

[0018] The dual-stage cascaded AI inference core processing module includes a first-stage real-time target detection and multi-target tracking unit, a second-stage spatiotemporal behavior recognition inference unit, and a hierarchical filtering unit for inference results. The output of the first-stage real-time target detection and multi-target tracking unit is connected to the input of the second-stage spatiotemporal behavior recognition inference unit; the output of the second-stage spatiotemporal behavior recognition inference unit is connected to the input of the hierarchical filtering unit for inference results. The first-stage real-time target detection and multi-target tracking unit completes multi-target detection and cross-frame trajectory association. The trajectory association uses a matching cost calculation method that fuses Mahalanobis distance and cosine similarity. The formula for calculating the fused matching cost is: , The cost of final fusion matching; This is the weighting coefficient, with a value ranging from 0.2 to 0.5; The Mahalanobis distance; The first stage is the cosine distance; the second stage, the spatiotemporal dimension behavior recognition reasoning unit, completes the temporal feature extraction and category recognition of the target behavior; the reasoning result hierarchical filtering unit completes the deduplication and filtering of the reasoning results; Specifically, the inference result hierarchical filtering unit: uses non-maximum suppression to remove duplicate detection boxes, sets hierarchical confidence thresholds to filter low-confidence results, combines scenario business rules to complete secondary verification, filters false detection results, and outputs accurate inference results; The second-stage spatiotemporal behavior recognition reasoning unit: Based on the target area locked by target detection and continuous temporal images, it extracts the spatial semantic features and action temporal features of the target to complete the accurate recognition of the preset behavior category; its implementation logic is to use a dual-branch SlowFast temporal behavior recognition network to separate spatial and temporal features, complete feature fusion through channel attention weighting, and finally output the behavior recognition result and corresponding confidence score. The first stage, real-time target detection and multi-target tracking unit, completes real-time detection of multiple targets and continuous trajectory tracking across frames, providing complete temporal trajectory data and target locking regions for subsequent behavior recognition. Its implementation logic is to use an optimized YOLO detection network to complete target detection, combined with the ByteTrack multi-target tracking algorithm, and calculate the optimal matching of trajectory and detection box through the matching cost fusion of Mahalanobis distance and cosine similarity. A unique tracking ID is assigned to each target to avoid trajectory breakage caused by target occlusion and rapid movement. The standard format for the core fusion matching cost calculation formula is as follows: The formula is used to fuse the spatial location matching degree and appearance feature matching degree of the target to achieve accurate association between the trajectory and the detection box, and solve the trajectory matching error problem caused by a single matching method; where d is the final fusion matching cost, which is the output value of this formula and is used for the optimal matching calculation of the Hungarian algorithm. The lower the value, the higher the matching degree. This is a weighting coefficient, ranging from 0.2 to 0.5, used to balance the weight ratio between spatial location matching degree and appearance feature matching degree; The Mahalanobis distance is used to measure the spatial matching degree between the predicted position of the target trajectory and the current detection box, and is calculated based on the trajectory prediction results of Kalman filtering. The cosine distance is used to measure the matching degree of the target's appearance features. It is calculated based on the appearance feature vector of the detected target and the historical appearance feature vector of the trajectory.

[0019] The intelligent event analysis and coordinated response module includes an event classification and rule matching unit, an event tracing and evidence retention unit, a multi-system coordinated response unit, and an event archiving and statistical analysis unit. The output of the event classification and rule matching unit is connected to the input of the event tracing and evidence retention unit; the output of the event tracing and evidence retention unit is connected to the input of the multi-system coordinated response unit; the output of the multi-system coordinated response unit is connected to the input of the event archiving and statistical analysis unit. The event classification and rule matching unit classifies events by risk level and matches response procedures; the event tracing and evidence retention unit stores corresponding videos and key information; the multi-system coordinated response unit handles coordinated response and information dissemination; and the event archiving and statistical analysis unit archives and statistically analyzes event data. Specifically, the event classification and rule matching unit: based on a scenario-based business rule library, it completes the event judgment, risk level classification and corresponding handling process matching of AI inference results; its implementation logic is to build a hierarchical business rule library, adopt a multi-condition weighted verification mechanism, and perform matching calculations on multi-dimensional conditions such as target category, behavior type, and spatial region to complete the accurate judgment and classification of events. Event tracing and evidence retention unit: Automatically captures complete video clips and key frame screenshots before and after the alarm event, associates all key information corresponding to the event, generates an unalterable event ledger and evidence package, and completes the storage of event data; Multi-system linkage response unit: It connects with third-party business systems to realize on-site linkage response to alarm events, and accurately pushes alarm information to the corresponding management personnel to complete the closed-loop response of events; Event archiving and statistical analysis unit: Archives completed events, generates multi-dimensional statistical reports, and completes the full lifecycle management of event data.

[0020] The cloud-side data management and model iteration optimization module includes a scenario-based dataset construction unit, a model incremental training and lightweight optimization unit, a model canary release and OTA upgrade unit, and a global system operation and maintenance and status monitoring unit. The output of the scenario-based dataset construction unit is connected to the input of the model incremental training and lightweight optimization unit; the output of the model incremental training and lightweight optimization unit is connected to the input of the model canary release and OTA upgrade unit; the output of the model canary release and OTA upgrade unit is connected to the input of the global system operation and maintenance and status monitoring unit. The scenario-based dataset construction unit completes the classification, storage, and label injection of samples into the library; the model incremental training and lightweight optimization unit completes the parameter fine-tuning and structural optimization of the model; the model canary release and OTA upgrade unit completes the version management and push update of the model; and the global system operation and maintenance and status monitoring unit completes the monitoring of the operational status of all system devices and nodes. Specifically, the model canary release and OTA upgrade unit: establishes a model version management system, completes the canary pilot verification and full push upgrade of the optimized model, and supports version rollback to ensure the stability of model updates at edge nodes; The global system operation and maintenance and status monitoring unit monitors the online status, computing power usage, storage capacity, running frame rate and other core parameters of the entire system's end-side devices and edge nodes in real time, enabling automatic early warning of equipment failures and centralized operation and maintenance; Scenario-based dataset construction unit: It gathers effective samples from the entire system to build a continuously updated scenario-based dataset, providing data support for model iteration and optimization. Its implementation logic is to gather alarm data, difficult sample, and false positive / false negative sample uploaded by edge nodes, complete classification and storage, and use a semi-automatic annotation method of model pre-annotation plus manual verification to complete the annotation and storage of samples, and build a dedicated dataset adapted to the current business scenario. The incremental training and lightweight optimization unit optimizes model accuracy based on scenario-based datasets while simultaneously lightweighting the model to adapt to the computational limitations of edge nodes. Its implementation logic involves freezing the weights of the backbone network based on newly added difficult examples, and only incrementally fine-tuning the detection and classification heads to avoid the computational cost of full retraining. Furthermore, it employs knowledge distillation, INT8 quantization, and structured pruning techniques to optimize model size and inference speed without sacrificing core accuracy, thus adapting to the deployment requirements of edge devices.

[0021] The method of use and working principle of this invention are as follows: Usage: First, complete the on-site deployment and point layout of various types of visual acquisition terminals. Connect and adapt each acquisition terminal to the system's multi-protocol compatible access unit. Through the cloud platform, complete the preset configuration of the business scenario's focus area, behavior recognition category, event classification rules, and linkage handling process. After configuration, start the system's full-process operation. During operation, the on-site video footage, event alarm information, and equipment operating status can be viewed in real time through the cloud platform. For alarm events, the platform can be used to complete the entire process of dispatching, evidence retrieval, and result feedback. At the same time, the cloud platform can be used to complete related operations such as system parameter adjustment, model version switching, and remote equipment maintenance, realizing the full lifecycle management of the system.

[0022] Working principle: After the multi-source visual data acquisition and protocol adaptation module completes the acquisition, protocol adaptation, and initial screening of invalid frames of the on-site video stream, the valid video stream is encrypted and transmitted to the edge node. The edge data standardization preprocessing and feature enhancement module completes the hardware decoding of the video stream, image format standardization, non-interest area filtering, and image quality enhancement. The processed image data is input into the dual-stage cascaded AI inference core processing module. The first stage completes the real-time detection of multiple targets and cross-frame trajectory tracking. The second stage completes the temporal feature extraction and accurate recognition of target behavior. The inference results are input into the intelligent event analysis and linkage response module to complete event classification, rule matching, evidence retention, linkage response, and data archiving. At the same time, the cloud-side data management and model iteration optimization module gathers the sample data of the entire system to build a scenario-based dataset, completes the incremental training and lightweight optimization of the model, and distributes the optimized model to the edge node for iterative updates. At the same time, it realizes the centralized operation and maintenance management of the entire system's equipment, forming a closed-loop operation logic of data acquisition, analysis and inference, business processing, and model iteration.

[0023] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the description and drawings above. However, any modifications, alterations, and variations made by those skilled in the art without departing from the scope of the present invention using the disclosed technical content are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, and variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. An AI-powered intelligent analysis system for visual target detection and behavior recognition, characterized in that, It includes a multi-source visual data acquisition and protocol adaptation module, a side-side data standardization preprocessing and feature enhancement module, a two-stage cascaded AI inference core processing module, an intelligent event analysis and coordinated response module, and a cloud-side data management and model iteration optimization module. The output of the multi-source visual data acquisition and protocol adaptation module is connected to the input of the side-side data standardization preprocessing and feature enhancement module; the output of the side-side data standardization preprocessing and feature enhancement module is connected to the input of the two-stage cascaded AI inference core processing module; the output of the two-stage cascaded AI inference core processing module is connected to the input of the intelligent event analysis and coordinated response module; the output of the intelligent event analysis and coordinated response module is connected to the input of the cloud-side data management and model iteration optimization module; and the output of the cloud-side data management and model iteration optimization module is connected to the input of the two-stage cascaded AI inference core processing module.

2. The AI ​​visual target detection and behavior recognition intelligent analysis system according to claim 1, characterized in that, The multi-source visual data acquisition and protocol adaptation module includes multiple types of visual acquisition terminal devices, a multi-protocol compatible access unit, an end-side dynamic frame initial screening unit, and an encrypted transmission and link keep-alive unit. The output of the multiple types of visual acquisition terminal devices is connected to the input of the multi-protocol compatible access unit; the output of the multi-protocol compatible access unit is connected to the input of the end-side dynamic frame initial screening unit; the output of the end-side dynamic frame initial screening unit is connected to the input of the encrypted transmission and link keep-alive unit; the end-side dynamic frame initial screening unit performs grayscale processing on continuously input video frames to obtain the first... A frame grayscale image is denoted as ;No. A frame grayscale image is denoted as ; Calculate the absolute frame difference between two adjacent frames. ; Calculate the global motion saliency value of the frame difference image. ; Construct adaptive threshold Complete the frame validity check; The formula for calculating the adaptive threshold T_t is: , This is a dynamic weighting coefficient; its value ranges from 0.6 to 0.

8. These are static weighting coefficients; their values ​​range from 0.2 to 0.

4. The preset base threshold; This represents the mean global motion saliency of the previous frame.

3. The AI ​​visual target detection and behavior recognition intelligent analysis system according to claim 1, characterized in that, The side-side data standardization preprocessing and feature enhancement module includes a hardware-accelerated decoding and format standardization unit, a Region of Interest (ROI) filtering unit, and an adaptive image quality enhancement unit. The output of the hardware-accelerated decoding and format standardization unit is connected to the input of the ROI filtering unit; the output of the ROI filtering unit is connected to the input of the adaptive image quality enhancement unit; the hardware-accelerated decoding and format standardization unit performs hardware decoding of the video stream and image format unification processing; the ROI filtering unit performs pixel-level filtering of non-interested regions. The adaptive image quality enhancement unit performs quality optimization and feature enhancement on the input image.

4. The AI ​​visual target detection and behavior recognition intelligent analysis system according to claim 1, characterized in that, The dual-stage cascaded AI inference core processing module includes a first-stage real-time target detection and multi-target tracking unit, a second-stage spatiotemporal dimension behavior recognition inference unit, and a inference result hierarchical filtering unit; the output of the first-stage real-time target detection and multi-target tracking unit is connected to the input of the second-stage spatiotemporal dimension behavior recognition inference unit. The output of the second-stage spatiotemporal behavior recognition inference unit is connected to the input of the inference result hierarchical filtering unit; the first-stage real-time target detection and multi-target tracking unit completes multi-target detection and cross-frame trajectory association; the trajectory association adopts a matching cost calculation method that fuses Mahalanobis distance and cosine similarity; the formula for calculating the fused matching cost is: , The cost of final fusion matching; This is the weighting coefficient, with a value ranging from 0.2 to 0.5; The Mahalanobis distance; Cosine distance; The second stage, the spatiotemporal dimension behavior recognition reasoning unit, completes the extraction of temporal features and category recognition of the target behavior; the reasoning result hierarchical filtering unit completes the deduplication and filtering of reasoning results.

5. The AI ​​visual target detection and behavior recognition intelligent analysis system according to claim 1, characterized in that, The intelligent event analysis and coordinated response module includes an event classification and rule matching unit, an event tracing and evidence retention unit, a multi-system coordinated response unit, and an event archiving and statistical analysis unit. The output of the event classification and rule matching unit is connected to the input of the event tracing and evidence retention unit; the output of the event tracing and evidence retention unit is connected to the input of the multi-system coordinated response unit; the output of the multi-system coordinated response unit is connected to the input of the event archiving and statistical analysis unit. The event classification and rule matching unit completes the risk level classification of events and the matching of response procedures; the event tracing and evidence retention unit completes the storage of corresponding videos and key information of events; the multi-system coordinated response unit completes the coordinated response and information push of events; and the event archiving and statistical analysis unit completes the archiving and statistics of event data.

6. The AI ​​visual target detection and behavior recognition intelligent analysis system according to claim 1, characterized in that, The cloud-side data management and model iteration optimization module includes a scenario-based dataset construction unit, a model incremental training and lightweight optimization unit, a model canary release and OTA upgrade unit, and a global system operation and maintenance and status monitoring unit. The output of the scenario-based dataset construction unit is connected to the input of the model incremental training and lightweight optimization unit; the output of the model incremental training and lightweight optimization unit is connected to the input of the model canary release and OTA upgrade unit; and the output of the model canary release and OTA upgrade unit is connected to the input of the global system operation and maintenance and status monitoring unit. The scenario-based dataset construction unit completes the classification, storage, and labeling of samples into the library; the incremental training and lightweight optimization unit completes the parameter fine-tuning and structural optimization of the model; and the model canary release and OTA upgrade unit completes the version management and push update of the model. The global system operation and maintenance and status monitoring unit completes the monitoring of the operational status of all system devices and nodes.

Images