Expressway video event accurate analysis method and system based on edge calculation

Abstract

The invention relates to the field of intelligent traffic and computer vision, and particularly discloses a highway video event accurate analysis method and system based on edge calculation. According to the method, an event detection model and a multi-dimensional context fusion research and judgment mechanism are deployed at a roadside edge computing node, so that real-time analysis and local precise processing of a video stream are realized; meanwhile, the system combines a rule engine and a micro model to carry out cooperative verification and intelligent decision making on events, and the accuracy and reliability of event identification and classification are effectively improved. According to the invention, a bidirectional learning closed loop of continuous iteration of local optimization feedback and a cloud model is established, so that the system can adapt to traffic scene changes and evolve continuously. The system effectively overcomes the problems of large network delay, high bandwidth pressure, low precision of a front-end light computing scheme and many false alarms of a traditional cloud processing mode.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and computer vision, specifically a method and system for accurate analysis of highway video events based on edge computing. Background Technology

[0002] As the arteries of modern transportation networks, the safety and efficiency of highways are of paramount importance. Video surveillance is a core means of acquiring road condition information and detecting events. With the development of artificial intelligence technology, automatic video event detection technology based on computer vision has become a research hotspot. It aims to replace or assist manual inspections, achieving real-time, automatic identification and alerting for events such as traffic accidents, abnormal parking, and littering.

[0003] Traditional video event detection solutions typically employ a centralized architecture of "front-end acquisition + cloud processing." In this model, high-definition cameras deployed on the roadside transmit massive video streams in real-time via networks (such as fiber optics or 5G) to remote cloud data centers or central servers. In the cloud, large-scale artificial intelligence algorithms with powerful computing capabilities but complex models analyze the video streams to identify potential events. The core drawback of this architecture lies in its processing lag and network dependency. Because high-definition video stream data transmission requires significant bandwidth and inevitably introduces network latency, there can be a delay of several seconds or even longer between the occurrence of an event and the completion of analysis and the issuance of an alarm in the cloud. This is unacceptable for highway safety scenarios requiring "second-level" or even "sub-second-level" response times (such as debris warnings). Furthermore, network fluctuations or interruptions will directly lead to monitoring failure, severely limiting the system's reliability and real-time performance.

[0004] To reduce latency, solutions have emerged in recent years that deploy algorithms inside or near cameras for simple motion detection or area intrusion detection. However, these methods often sacrifice model accuracy and analytical depth due to limited device computing resources. They can often only perform preliminary target discovery or simple rule judgments, failing to combine historical data, real-time traffic flow, road conditions, and other multi-dimensional contextual information to accurately and reliably determine the type, severity, and authenticity of events. This results in high false alarm and false negative rates, limiting their practical application value and still requiring frequent manual intervention for verification. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for precise analysis of highway video events based on edge computing. This addresses the issues in existing technologies where "centralized computing" leads to insufficient real-time performance, and "light front-end computing" results in limited accuracy and intelligence, making it difficult to achieve rapid, accurate, and reliable event perception in resource-constrained on-site environments.

[0006] A method for accurate analysis of highway video events based on edge computing, deployed on roadside edge computing nodes, includes the following steps:

[0007] S1: Video event detection and preliminary analysis, specifically:

[0008] The video acquisition equipment is used to acquire real-time video streams of the monitored area, and the video streams are decoded and preprocessed at the edge computing node to extract spatiotemporal features;

[0009] Based on the extracted spatiotemporal features, the event detection model is run to generate preliminary event alerts containing event type, location, and confidence level.

[0010] S2: Event Correlation Analysis and Precise Reporting, specifically:

[0011] Receive the initial event alert and obtain the multidimensional context information associated with the alert;

[0012] Based on a fusion rule adapted to the spatiotemporal feature extraction algorithm in step S1, the feature information of the preliminary event alarm is fused and analyzed with the multidimensional context information to accurately judge the event and generate an accurate event report.

[0013] The accurate event report is uploaded to the cloud control center.

[0014] Preferably, the step of decoding and preprocessing the video stream at the edge computing node to extract spatiotemporal features specifically includes:

[0015] S11: Decode and align the input video stream to obtain a preprocessed video frame sequence;

[0016] S12: Perform multi-scale convolution operations on the preprocessed video frame sequence to extract spatial detail features and semantic features;

[0017] S13: Perform temporal modeling on a continuous video frame sequence to extract motion change features;

[0018] S14: Integrate the spatial detail features, semantic features, and motion change features to generate a spatiotemporal feature vector for event detection.

[0019] Preferably, the fusion analysis based on a fusion rule adapted to the spatiotemporal feature extraction algorithm in step S1 specifically includes:

[0020] S21: Receive the preliminary event alarm and its corresponding spatiotemporal feature vector;

[0021] S22: Based on the event location and timestamp, retrieve relevant historical event features, real-time traffic flow data and road status information from the local cache of the edge node to form a multi-dimensional context information vector;

[0022] S23: Based on the weight distribution of the features extracted in steps S12 and S13 in the spatiotemporal feature vector, dynamically calculate the fusion coefficient between the current alarm feature and the context information vector;

[0023] S24: Use the fusion coefficient to perform weighted fusion of the two types of vectors to generate an enhanced event feature vector for subsequent accurate analysis.

[0024] Preferably, the precise analysis of the event specifically includes:

[0025] The enhanced event feature vector is input into a judgment module that combines a rule engine and a micro-model;

[0026] The rule engine performs logical validation on event types based on pre-defined business rules;

[0027] The micro-model is trained based on historical data and fine-tuned for event level, impact range, and confidence level;

[0028] The outputs of the combined rules engine and micro-model are used to generate the accurate event report, which includes the final event type, level, location, confidence level, and handling recommendations.

[0029] Preferably, the event detection model is obtained by compressing a large-scale detection model deployed in the cloud control center through knowledge distillation technology, and periodically receives incremental update parameters from the cloud control center to adapt to newly emerging event patterns.

[0030] Preferably, after generating accurate event reports, a local feedback optimization step is also included:

[0031] Key features and analysis results from accurate event reports are stored in the local event database of edge nodes;

[0032] When a similar event occurs again, the optimized features stored in the local event database are used first for matching and analysis to improve the speed of detection and analysis of subsequent similar events.

[0033] A highway video event precision analysis system based on edge computing, used to implement the method described above, includes: a cloud control center, multiple edge computing nodes, and video acquisition equipment deployed along the highway;

[0034] The video acquisition device is used to acquire real-time video streams of the monitored area and send them to the corresponding edge computing nodes;

[0035] The edge computing node includes at least:

[0036] The video processing and feature extraction module is configured to perform the aforementioned video decoding, preprocessing, and spatiotemporal feature extraction functions.

[0037] The event detection module is configured to run the event detection model and generate initial event alerts.

[0038] The context fusion and precise judgment module is configured to perform the above-mentioned correlation analysis, feature fusion and precise judgment functions. The feature fusion unit contained therein is adapted to the algorithm of the video processing and feature extraction module and interacts with the data.

[0039] The data reporting module is configured to upload accurate event reports to the cloud control center.

[0040] The cloud control center is used to receive and aggregate event reports from various edge computing nodes, and to perform global situational awareness, model training, and policy distribution.

[0041] Preferably, the edge computing node further includes a local cache and optimization library for storing historical event features, multi-dimensional context information, and optimized features and judgment rules accumulated through local feedback optimization steps.

[0042] Preferably, the cloud control center includes a global event library, a model training platform, and a policy management engine, used for unified management of all edge computing nodes, model iteration, and the updating and distribution of judgment policies.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] By deploying models on roadside edge computing nodes for real-time detection and preliminary assessment, sub-second response and localized closed-loop processing of events are achieved, effectively overcoming the network latency and bandwidth pressure caused by traditional centralized processing. This ensures early detection and warning of emergencies such as spilled objects, illegal parking, and traffic accidents, and significantly improves the proactive safety capabilities of highway operations.

[0045] By dynamically adjusting the fusion weights based on the semantics and motion saliency of the video features themselves, and combining the logical verification of the rule engine with the empirical tuning of the micro-model, it can effectively filter out false alarms caused by interference factors such as changes in lighting and vehicle shadows. At the same time, it can make judgments on the event type, level and impact range in complex scenes that are closer to the level of human experts, thus reducing the false alarm rate and the missed alarm rate.

[0046] By establishing a two-way learning loop of local feedback optimization and centralized cloud iteration, the system possesses continuous evolution and adaptive capabilities. Edge nodes locally store the features of successfully processed cases, forming an "experience base" to accelerate the processing of subsequent similar events; the cloud aggregates global data, continuously optimizes the core model, and distributes updates. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] Reference Figure 1 This is the first embodiment of the present invention, providing a method for accurate analysis of highway video events based on edge computing, comprising the following steps:

[0051] S1: Video event detection and preliminary analysis.

[0052] Specifically, video event detection and preliminary assessment are achieved by processing real-time video streams at roadside edge computing nodes to quickly generate preliminary event alerts. The specific implementation is as follows:

[0053] Real-time video streams of highway monitoring areas are collected using video capture equipment. ,in Represents a timestamp;

[0054] Based on the acquired real-time video stream, decoding and preprocessing are performed at edge computing nodes to extract spatiotemporal features, specifically:

[0055] The input video stream is decoded and its frames are aligned to obtain a preprocessed video frame sequence. ,in Indicates the length of the processing time window;

[0056] Multi-scale convolution operations are performed on the preprocessed video frame sequence to extract spatial detail and semantic features. Simultaneously, temporal modeling is performed on the continuous video frame sequence to extract motion change features. Then, we have...

[0057] Using the first convolutional neural network Extracting detailed texture features from video frames :

[0058]

[0059] Using the second convolutional neural network Extracting macroscopic semantic features from pooled video frames :

[0060]

[0061] Using time series modeling networks (e.g., LSTM or GRU) to analyze motion change characteristics between consecutive video frames :

[0062]

[0063] in, This indicates the size of the timing window, which can be set by the implementer based on the actual application scenario.

[0064] By fusing detailed texture features, semantic features, and motion variation features, a spatiotemporal feature vector for event detection is generated. :

[0065]

[0066] in, , , These represent the weight coefficients of the corresponding features. This indicates a vector concatenation operation.

[0067] Based on the extracted spatiotemporal feature vectors Run the event detection model deployed on edge nodes Generate an event type ,Location and confidence level Preliminary incident alert :

[0068]

[0069] It should be noted that the event detection model deployed on the edge nodes... As an efficient model, its initial version utilizes knowledge distillation technology to extract knowledge from a large-scale, high-precision teacher model deployed in a cloud control center. This transfer approach significantly reduces model complexity and computational overhead while maintaining high detection performance. To adapt to constantly changing traffic scenarios and emerging event patterns, the cloud control center continuously trains the teacher model or generates model update knowledge packages (such as lightweight parameter increments) based on massive aggregated data, and distributes them to edge nodes via the network. Upon receiving the data, the edge nodes can integrate the updated knowledge into their local models using gradient-based optimization algorithms (such as stochastic gradient descent with momentum). The specific update process can be characterized as follows:

[0070]

[0071] in, The parameters representing the edge model, This indicates the updated knowledge being distributed (such as parameter difference or distillation loss gradient). For learning rate, This represents a specific parameter update function.

[0072] S2: Event Correlation Analysis and Precise Reporting.

[0073] Specifically, event correlation analysis and precise reporting are based on initial event alerts, integrating multi-dimensional contextual information for precise judgment, and reporting the results. The specific implementation is as follows:

[0074] Receive initial event alerts and its corresponding spatiotemporal feature vectors ;

[0075] Based on the location of the event in the alarm With timestamp Retrieve relevant historical event features from the local cache of edge nodes. Real-time traffic flow data and road condition information This constitutes a multidimensional context information vector. :

[0076]

[0077] in, Indicates the time range of historical searches.

[0078] Based on a fusion rule adapted to the spatiotemporal feature extraction algorithm in step S1, the feature information of the initial event alarm is fused and analyzed with multidimensional context information, specifically as follows:

[0079] Based on the detailed features, semantic features, and motion features extracted in step S1, in the spatiotemporal feature vector Weight distribution in , , The fusion coefficient between the current alarm features and the context information vector is dynamically calculated. :

[0080]

[0081] in, This represents the Sigmoid activation function. Let L2 norm be denoted. This formula implies that the higher the semantic saliency or motion intensity of an event, the more likely the fusion process is to trust the current alert features.

[0082] Using fusion coefficient The two types of vectors are weighted and fused to generate an enhanced event feature vector. :

[0083]

[0084] in, A learnable projection matrix used to map context information vectors to... Same feature space, It is the hyperbolic tangent activation function.

[0085] To conduct a precise analysis of the event, specifically:

[0086] Enhance event feature vectors Input a judgment module that integrates a rule engine and a micro-model;

[0087] The rules engine performs logical validation on event types based on pre-defined business rules (such as "parking event" must last for N frames, "spilled material" must be separated from moving vehicles);

[0088] micromodel Training based on historical data to determine event levels Scope of impact and confidence level Make minor adjustments:

[0089]

[0090] By combining the logical validation results of the rule engine with the output of the micro-model, accurate event reports are generated. The report includes the final event type, level, location, scope of impact, confidence level, and handling recommendations.

[0091] Accurate event reporting Uploaded to the cloud control center.

[0092] Furthermore, after generating accurate event reports, local feedback optimization steps are also included:

[0093] Key features in accurate event reports and the results of the analysis Local event libraries stored on edge nodes;

[0094] When a similar event occurs again, the optimized features stored in the local event database are used first for matching and analysis. If the similarity exceeds a threshold, the matching will be considered invalid. Then, it can directly call historical analysis results or perform rapid reasoning to improve the speed of detection and analysis of similar events in the future.

[0095] Example 2

[0096] A second embodiment of the present invention provides a highway video event precision analysis system based on edge computing, comprising:

[0097] Specifically, video acquisition equipment is deployed along the highway to collect real-time video streams of the monitored area and send them to the corresponding edge computing nodes;

[0098] Edge computing nodes, as core processing units, include at least:

[0099] Video processing and feature extraction module: This module receives the video stream, performs video decoding, frame alignment, multi-scale convolution, and temporal modeling operations as described in Example 1 (S1), and outputs high-quality spatiotemporal feature vectors. The performance of this module directly determines the input quality for subsequent event detection.

[0100] Event detection module: Loads and runs the lightweight model obtained through knowledge distillation, receives spatiotemporal feature vectors, and quickly generates preliminary event alerts. This module ensures real-time performance.

[0101] Context fusion and precise judgment module: This is the core intelligent unit of the system. Its internal feature fusion unit is adapted to the algorithms of the video processing and feature extraction modules, and can dynamically calculate fusion coefficients based on feature weights. This module receives initial alerts and multi-dimensional contextual information retrieved from the local cache. Through collaboration between the rule engine and the micro-model, it completes precise judgment of the event and generates a final report.

[0102] Data reporting module: responsible for packaging accurate event reports and uploading them to the cloud control center.

[0103] Local caching and optimization library: used to store historical event features, real-time / historical traffic flow, weather and other contextual information, as well as optimized features and judgment rules accumulated through local feedback optimization steps, providing data support for context fusion and rapid matching.

[0104] The cloud control center includes:

[0105] Global Event Repository: Aggregates event reports from all edge nodes to form a global situational map.

[0106] Model training platform: Based on massive amounts of data from a global event library, it trains and updates large-scale detection models. And generate models through knowledge distillation and its incremental update parameters It will be issued regularly.

[0107] Policy Management Engine: Unifies the management and distribution of business rules and micro-model iterations from the rule engines of each edge node, enabling centralized control and flexible configuration of policies.

[0108] System Workflow: Video capture devices continuously capture video and send it to edge nodes. The video processing and feature extraction module performs real-time feature calculations, and the event detection module generates an initial alert. The context fusion and precise judgment module immediately activates, retrieving relevant information from local caches and optimization libraries for deep fusion and intelligent judgment, generating a precise report. This report is sent to the cloud control center via the data reporting module for global awareness and model iteration; simultaneously, its key information is stored in the local optimization library for accelerated processing of similar future events. The cloud control center acts as the system's "brain," optimizing models and uniformly distributing strategies, forming a collaborative closed loop of "real-time edge processing + centralized cloud optimization."

[0109] The embodiments of the present invention are given for the purposes of illustration and description. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Any changes, modifications, substitutions and variations made by those skilled in the art to the above embodiments within the scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for accurate analysis of highway video events based on edge computing, characterized in that, Deployed on roadside edge computing nodes for execution, including the following steps: S1: Video event detection and preliminary analysis, specifically: The video acquisition equipment is used to acquire real-time video streams of the monitored area. The video streams are then decoded and preprocessed at the edge computing nodes to extract spatiotemporal features. Based on the extracted spatiotemporal features, the event detection model is run to generate preliminary event alerts containing event type, location, and confidence level. S2: Event Correlation Analysis and Precise Reporting, specifically: Receive initial event alerts and obtain multidimensional context information associated with the alert; Based on a fusion rule adapted to the spatiotemporal feature extraction algorithm in step S1, the feature information of the initial event alarm is fused and analyzed with multi-dimensional context information to accurately judge the event and generate an accurate event report. Upload accurate event reports to the cloud control center.

2. The method for accurate analysis of highway video events based on edge computing according to claim 1, characterized in that, The decoding and preprocessing of the video stream at the edge computing node to extract spatiotemporal features specifically includes: S11: Decode and align the input video stream to obtain a preprocessed video frame sequence; S12: Perform multi-scale convolution operations on the preprocessed video frame sequence to extract spatial detail features and semantic features; S13: Perform temporal modeling on a continuous video frame sequence to extract motion change features; S14: Integrate the spatial detail features, semantic features, and motion change features to generate a spatiotemporal feature vector for event detection.

3. The method for accurate analysis of highway video events based on edge computing according to claim 2, characterized in that, The fusion analysis is based on a fusion rule adapted to the spatiotemporal feature extraction algorithm in step S1, specifically as follows: S21: Receive the preliminary event alarm and its corresponding spatiotemporal feature vector; S22: Based on the event location and timestamp, retrieve relevant historical event features, real-time traffic flow data and road status information from the local cache of the edge node to form a multi-dimensional context information vector; S23: Based on the weight distribution of the features extracted in steps S12 and S13 in the spatiotemporal feature vector, dynamically calculate the fusion coefficient between the current alarm feature and the context information vector; S24: Use the fusion coefficient to perform weighted fusion of the two types of vectors to generate an enhanced event feature vector for subsequent accurate analysis.

4. The method for accurate analysis of highway video events based on edge computing according to claim 3, characterized in that, The precise analysis of events specifically includes: The enhanced event feature vector is input into a judgment module that combines a rule engine and a micro-model; The rule engine performs logical validation on event types based on pre-defined business rules; The micro-model is trained based on historical data and fine-tuned for event level, impact range, and confidence level; The outputs of the combined rules engine and micro-model are used to generate the accurate event report, which includes the final event type, level, location, confidence level, and handling recommendations.

5. The method for accurate analysis of highway video events based on edge computing according to claim 4, characterized in that, The event detection model is obtained by compressing a large-scale detection model deployed in the cloud control center using knowledge distillation technology, and periodically receives incremental parameter updates from the cloud control center to adapt to newly emerging event patterns.

6. The method for accurate analysis of highway video events based on edge computing according to claim 1, characterized in that, After generating accurate event reports, local feedback optimization steps are also included: Key features and analysis results from accurate event reports are stored in the local event database of edge nodes; When a similar event occurs again, the optimized features stored in the local event database are used first for matching and analysis to improve the speed of detection and analysis of subsequent similar events.

7. A highway video event precision analysis system based on edge computing, used to implement the method as described in any one of claims 1 to 6, characterized in that, include: A cloud control center, multiple edge computing nodes, and video acquisition equipment deployed along the highway; The video acquisition device is used to acquire real-time video streams of the monitored area and send them to the corresponding edge computing nodes; The edge computing node includes at least: The video processing and feature extraction module is configured to perform the video decoding, preprocessing, and spatiotemporal feature extraction functions as described in claim 1 or 2. The event detection module is configured to run the event detection model and generate initial event alerts. The context fusion and precise judgment module is configured to perform the association analysis, feature fusion and precise judgment functions as described in claim 3 or 4, and the feature fusion unit contained therein is adapted to the algorithm of the video processing and feature extraction module and interacts with data. The data reporting module is configured to upload accurate event reports to the cloud control center. The cloud control center is used to receive and aggregate event reports from various edge computing nodes, and to perform global situational awareness, model training, and policy distribution.

8. The highway video event precision analysis system based on edge computing according to claim 7, characterized in that, The edge computing node also includes a local cache and optimization library, used to store historical event features, multi-dimensional context information, and optimized features and judgment rules accumulated through local feedback optimization steps.

9. The highway video event precision analysis system based on edge computing according to claim 7, characterized in that, The cloud control center includes a global event library, a model training platform, and a policy management engine, which are used for unified management of all edge computing nodes, model iteration, and the updating and distribution of judgment policies.

Images