Search

CN-122024157-A - Real-time monitoring analysis system based on video intelligent analysis

CN122024157ACN 122024157 ACN122024157 ACN 122024157ACN-122024157-A

Abstract

The invention discloses a real-time monitoring analysis system based on intelligent video analysis, which relates to the technical field of intelligent video monitoring, wherein in the operation of the system, video information and environmental parameters in a monitoring area are synchronously collected, a continuous video stream is obtained based on a high dynamic range image sensor, and time information and space position information are embedded in a video frame to form a data source with space-time attribute; the method comprises the steps of preprocessing video streams in real time, extracting features, completing noise suppression, background modeling and target extraction, performing cross-view correlation and track fusion on multi-node target features, realizing space-time alignment analysis, sensing scene and metadata changes, dynamically adjusting analysis parameters, performing model online updating, performing semantic analysis and state judgment on the fused features, outputting hierarchical alarm information, intensively storing operation data, optimizing and retraining the model based on historical data, and synchronizing to an edge side to improve the overall performance of the system.

Inventors

  • ZHANG HAIYAN

Assignees

  • 千城云科(珠海横琴)科技有限公司

Dates

Publication Date
20260512
Application Date
20260130

Claims (10)

  1. 1. The real-time monitoring analysis system based on the video intelligent analysis is characterized by comprising a multidimensional sensing acquisition module, an edge heterogeneous calculation module, a space-time cooperative analysis module, a self-adaptive evolution control module, a semantic logic decision module and a cloud storage optimization module; The multi-dimensional sensing acquisition module is used for synchronously acquiring video information and environmental parameters in a monitoring area, acquiring continuous video streams based on a high dynamic range image sensor, and embedding high-precision time information and spatial position information in video frames to form a video data source with space-time attributes; The edge heterogeneous computing module is used for carrying out real-time preprocessing and feature extraction on the acquired video stream, completing noise suppression, background modeling and target region extraction through a parallel computing architecture, and generating a localized visual feature vector for subsequent analysis; the space-time collaborative analysis module is used for performing cross-view correlation and track integration on target features from different acquisition nodes, and realizing space-time alignment and fusion analysis of multi-source features by constructing dynamic topological relations and executing unified measurement space mapping; The self-adaptive evolution control module is used for sensing the change states of the video scene and the additional metadata thereof, dynamically adjusting calculation and analysis parameters according to the environment evolution condition, and executing online incremental update on the related model so as to maintain the stability and analysis precision of the system; the semantic logic decision module is used for carrying out behavior semantic analysis on the fused feature stream, judging the target state by combining probability reasoning and sequential logic, and outputting alarm information of corresponding level when the risk condition is met; the cloud storage optimization module is used for intensively storing system operation data and historical analysis results, optimizing and retraining the core algorithm model based on data accumulated for a long time, and synchronizing the optimized model to the edge side module through a safety mechanism so as to improve the overall performance.
  2. 2. The video intelligent analysis-based real-time monitoring analysis system according to claim 1, wherein the multi-dimensional sensing acquisition module further comprises an environmental monitoring array; the environment monitoring array comprises a photoresistor sensor, a high-precision humidity sensor and a thermistor temperature sensor, and is used for sampling environment parameters with 100 milliseconds as a period, carrying out logic alignment with a corresponding video frame sequence and packaging the environment parameters into a multidimensional sensing data packet comprising a video stream, a time stream, a position stream and an environment parameter stream; the multidimensional sensing acquisition module adopts a low-voltage differential signal transmission mechanism to send the multidimensional sensing data packet to the edge heterogeneous calculation module.
  3. 3. The video intelligent analysis-based real-time monitoring analysis system according to claim 1, wherein the processing logic inside the edge heterogeneous computing module comprises: carrying out weighted average calculation on each input frame of image by using a 3-by-3 pixel window so as to filter out thermal noise and high-frequency electronic interference of a sensor; Maintaining 3 to 5 Gaussian distribution models for each pixel point, and dynamically updating the weight, the mean value and the variance of each distribution through an online expectation maximization algorithm; When the deviation between the newly input pixel value and the existing Gaussian distribution exceeds 2.5 times of standard deviation, judging the pixel as a motion prospect; and performing morphological closing operation on the extracted motion foreground region, filling the cavity in the target and removing isolated noise blocks to lock the target candidate region.
  4. 4. The video intelligent analysis-based real-time monitoring analysis system according to claim 1, wherein the process of performing cross-camera correlation by the spatio-temporal collaborative analysis module comprises: extracting target visual characterization vectors captured by two monitoring nodes, and calculating cosine similarity of the two target visual characterization vectors; Acquiring actual time intervals of disappearance and appearance of a target between two monitoring nodes, and extracting preset average transfer time in a camera topological distribution model; acquiring a first speed vector direction when the target disappears and a second speed vector direction when the target appears again; And carrying out weighted summation on the cosine similarity, an exponential decay term constructed by the difference between the actual time interval and the average transfer time and cosine values of the first speed vector direction and the second speed vector direction to obtain a matching score, and judging the matching score to be the same target and automatically merging the motion tracks when the matching score exceeds 0.85.
  5. 5. The system for real-time monitoring and analyzing based on intelligent video analysis according to claim 1, wherein the space-time collaborative analysis module is further integrated with a virtual 3D scene reconstruction function; The virtual 3D scene reconstruction function restores real-time three-dimensional dynamic pictures of a monitoring site in a virtual space by utilizing video streams with a plurality of different angles through a multi-view geometric calibration and texture mapping technology; the virtual 3D scene reconstruction function supports a manager to enter a monitoring scene through virtual reality equipment at a first view angle, so that immersive patrol and command scheduling are realized.
  6. 6. The video intelligent analysis-based real-time monitoring analysis system according to claim 1, wherein the operation mechanism of the adaptive evolution control module comprises: collecting the signal-to-noise ratio and the average brightness distribution parameter returned by the front end in real time; when the illumination intensity is detected to be lower than 10 lux or the coverage area ratio is detected to be more than 40%, the edge heterogeneous computation module is instructed to switch to a near infrared enhancement operator or an image restoration model based on a generated countermeasure network is adopted; Executing an online knowledge distillation strategy, performing fine adjustment on a local student model by using a teacher model issued by the cloud storage optimization module, and realizing online incremental update on a specific illumination environment or a visual angle inclination angle through a labeled sample.
  7. 7. The video intelligent analysis-based real-time monitoring analysis system according to claim 1, wherein the semantic logic decision module performs a behavior decision process comprising: acquiring a target motion characteristic and an interaction relation of more than 50 continuous frames, and inputting the extracted space coordinate sequence into a long-period memory neural network to learn a behavior time sequence evolution characteristic of the target; constructing a prior probability library containing normal behavior norms, and calculating posterior probability of the abnormal norms of the current observation sequence through Bayesian inference; The abnormal paradigm comprises illegal invasion, personnel gathering, reverse driving and object leaving, and when the posterior probability is higher than 0.9, decision execution logic is triggered immediately.
  8. 8. The video intelligent analysis-based real-time monitoring analysis system according to claim 7, wherein the decision execution logic comprises a multi-stage feedback mechanism: Aiming at low-risk suspected abnormalities with the abnormal probability of 0.7-0.9, generating prompting information and pushing the prompting information to an on-duty terminal; Marking the original video data in the abnormal time period as high priority, and triggering a cloud backup process to ensure the integrity of the evidence chain.
  9. 9. The video intelligent analysis-based real-time monitoring analysis system according to claim 1, wherein the cloud storage optimization module further comprises a model compression engine; The model compression engine iteratively updates the deep neural network by utilizing characteristic data converged from a plurality of monitoring nodes; and the model compression engine carries out quantization and pruning treatment on the retrained model, converts the 32-bit floating point number model into an 8-bit fixed point number model, eliminates neuron connection with contribution degree lower than 5%, and distributes the optimized model operator to the edge heterogeneous calculation module.
  10. 10. The video intelligent analysis-based real-time monitoring analysis system according to claim 1, further comprising a security architecture and a dynamic bandwidth adaptation protocol; The security protection architecture requires the multidimensional sensing acquisition module to carry out digital signature based on hardware trust root on all acquired data packets, and adopts a symmetric encryption algorithm to encrypt data streams in the transmission process; The safety protection architecture records the operation behaviors of the semantic logic decision module by using a blockchain technology to form an audit log; The dynamic bandwidth self-adaptive protocol adjusts the data transmission priority in real time according to the quality of the communication link, and preferentially transmits the semantic attribute and the alarm metadata of the target when the network bandwidth is limited.

Description

Real-time monitoring analysis system based on video intelligent analysis Technical Field The invention relates to the technical field of intelligent video monitoring, in particular to a real-time monitoring analysis system based on intelligent video analysis. Background With the rapid development of computer vision and deep learning technology, intelligent video analysis has become a core support for constructing modern smart cities and industrial Internet. In real-time monitoring scenes such as public safety, industrial inspection and complex traffic control, the technology greatly improves the efficiency and depth of information processing by automatically analyzing and extracting features of massive video data. The real-time monitoring analysis system is not only an important infrastructure for digital transformation in various industries, but also a key technical means for realizing high-precision environment sensing, risk avoidance and intelligent decision support. The real-time monitoring technology based on intelligent video analysis focuses on integrating a deep neural network model into a monitoring business process so as to realize automatic capturing and dynamic tracking of specific targets in a monitoring area. The core of the technical direction is to analyze the input original video stream in real time by utilizing an efficient algorithm framework, so as to identify the target attribute and judge the complex behavior mode. An ideal system can maintain stable perception performance under changeable environmental conditions, trigger an early warning feedback mechanism of millisecond level aiming at abnormal events, and ensure continuity, accuracy and initiative of monitoring tasks. However, the existing video monitoring technology still faces multiple bottlenecks in practical application, and the traditional system excessively depends on manual on duty or preset simple judgment rules, so that multi-objective behavior semantics are difficult to understand deeply when complex dynamic scenes are processed, and the problems of high false alarm rate and response lag are often caused. Meanwhile, the mainstream analysis system mostly adopts an offline training and static deployment mode, and when the model is in face of actual environment disturbance such as severe illumination change, frequent object shielding or monitoring visual angle migration, the model generalization capability is insufficient and the recognition accuracy is remarkably attenuated. In addition, the video stream space-time alignment and semantic fusion mechanism under the multi-camera collaborative monitoring scene is deficient, so that effective global situation awareness cannot be formed, and the analysis task is excessively concentrated on a back-end server, so that the network bandwidth pressure is increased rapidly, and the real-time and high reliability of monitoring are difficult to ensure under the weak network or extremely delay sensitive scene. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a real-time monitoring analysis system based on intelligent video analysis, which solves the problems in the background art. The real-time monitoring analysis system based on the video intelligent analysis comprises a multidimensional sensing acquisition module, an edge heterogeneous calculation module, a space-time cooperative analysis module, a self-adaptive evolution control module, a semantic logic decision module and a cloud storage optimization module; The multi-dimensional sensing acquisition module is used for executing real-time acquisition of a high-resolution video stream and synchronization of multi-mode environment parameters, the multi-dimensional sensing front end unit is deployed at a physical node of an area to be monitored, is provided with a high dynamic range image sensor and an embedded synchronous clock, performs gain control and exposure compensation through an image signal processor, converts an original optical signal into a digital video signal with a frame per second of not less than 30, and embeds a nanosecond time stamp and a geospatial coordinate attribute acquired by an integrated positioning chip in a metadata area of the video frame; The edge heterogeneous computing module is used for receiving an original video stream, carrying out pixel-level preprocessing, target area detection and preliminary feature extraction, adopting a pipeline parallel processing architecture, denoising by using a spatial domain low-pass filter realized by hard kernel logic, carrying out background modeling based on an improved Gaussian mixture model to extract a moving target candidate area, and calling a convolution operator in a neural network processor to extract a 128-dimensional fixed-point visual representation vector of a target; The space-time collaborative analysis module is used for executing target association, track completion and global situation awareness of a view field cr