Highway event detection system and method based on multivariate algorithm model

Abstract

The invention relates to the technical field of intelligent traffic management, and discloses an expressway event detection system and method based on a multivariate algorithm model, and the system comprises a sensing equipment access layer which is used for accessing and collecting the video stream data of various types of cameras deployed along an expressway; the multivariate algorithm analysis layer is connected to the sensing equipment access layer and is used for receiving the video stream data and calling a plurality of special event detection algorithm models to carry out parallel or selective real-time analysis on the video stream data so as to detect traffic events, analyze traffic situations and monitor equipment states; and the comprehensive application display layer is connected to the multivariate algorithm analysis layer and is used for receiving the analysis result and carrying out visual display and alarm. The method corresponds to the system. According to the invention, the problem of information islands formed by single detection capability and isolated functional modules of an existing system is solved, comprehensive, synchronous and automatic detection, research and judgment of various types of expressway events are realized, and the comprehensive management and control efficiency of a road network is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent traffic management technology, specifically a highway event detection system and method based on a multivariate algorithm model. Background Technology

[0002] With the widespread adoption of video surveillance facilities on highways, utilizing video analytics for automatic event detection has become a key means of improving road network management efficiency. However, existing highway event detection systems and methods typically have significant limitations. Most systems rely on single or a few algorithm models, capable of identifying only specific types of events such as parking and wrong-way driving. They struggle to achieve comprehensive coverage and simultaneous monitoring of a wide range of complex events, including traffic accidents, road debris, illegal construction, pedestrian intrusion, non-motorized vehicle traffic, and traffic congestion. Furthermore, even when multiple detection functions are deployed, they often operate independently, forming information silos without coordination. The lack of a unified framework for collaborative analysis and information fusion of multi-source video streams results in insufficient real-time performance, accuracy, and comprehensive judgment capabilities in event detection, failing to provide comprehensive, reliable, and integrated intelligent decision support for highway operation and management. Summary of the Invention

[0003] The purpose of this application is to provide a highway event detection system and method based on a multivariate algorithm model to solve the technical problems mentioned in the background.

[0004] To achieve the above objectives, this application discloses the following technical solutions: In a first aspect, this application discloses a highway event detection system based on a multivariate algorithm model, the system comprising: The sensing device access layer is used to access and collect video stream data from various types of cameras deployed along the highway; The multi-algorithm analysis layer, connected to the sensing device access layer, is used to receive the video stream data and call multiple pre-set dedicated event detection algorithm models to perform parallel or selective real-time analysis of the video stream data in order to detect traffic events, analyze traffic conditions, and monitor device status. The integrated application display layer is connected to the multi-element algorithm analysis layer, and is used to receive the analysis results of the multi-element algorithm analysis layer, and to visualize and alarm the traffic events, traffic conditions and equipment status.

[0005] Optionally, the multivariate algorithm analysis layer includes: A video analytics server cluster, wherein the video analytics server cluster is equipped with the multiple dedicated event detection algorithm models; Edge computing devices, deployed along highways, are used to preprocess and perform preliminary analysis on video stream data from the access layer of the nearby sensing devices, and transmit the analysis results or video feature data to the video analysis server cluster.

[0006] Optionally, the video analytics server cluster includes servers that employ an x86 architecture and are configured with NVIDIA GPU accelerator cards or Cambricon smart accelerator cards.

[0007] Optionally, the multiple dedicated event detection algorithm models include: pedestrian detection model, parking event detection model, congestion detection model, construction event detection model, non-motorized vehicle detection model, wrong-way driving detection model, litter detection model, vehicle speeding detection model, and vehicle slow-moving detection model.

[0008] Optionally, the pedestrian detection model is used to identify pedestrian targets entering the highway surface within a set detection area, and to generate pedestrian intrusion alarm information when the duration of the pedestrian target's existence reaches a first preset threshold. The parking event detection model is used to identify vehicle targets that change from a driving state to a stopped state in a highway lane, and to generate parking event alarm information when the stopping time of the vehicle target exceeds a second preset threshold.

[0009] Optionally, the construction event detection model is used to simultaneously identify traffic cones, construction vehicles, and workers within a set detection area, and to generate construction event alarm information when the traffic cones, construction vehicles, and workers are present at the same time. The spill detection model is used to identify spilled object targets with a size not less than a preset pixel size threshold within a set detection area, and to generate a spill event alarm message when the continuous detection time of the spilled object target at the same position reaches a third preset threshold.

[0010] Optionally, the sensing device access layer supports access to cameras from different manufacturers via RTSP, ONVIF, or GB28181 protocols; the types of cameras include PTZ cameras, dome cameras, bullet cameras, fisheye cameras, and integrated camera capture devices.

[0011] Optionally, the multi-element algorithm analysis layer also includes an alarm logic optimization module; The alarm logic optimization module is used to filter or classify the alarm information generated by the multiple dedicated event detection algorithm models; wherein, it suppresses repeated alarm information triggered by the same traffic event, and / or classifies pedestrian intrusion alarm information according to whether the identified person is wearing work clothes with preset characteristics.

[0012] Optionally, the multivariate algorithm analysis layer further includes an online calibration and target spatiotemporal fusion module; The online calibration and target spatiotemporal fusion module is used to calibrate the camera and construct a mapping model between the image coordinate system and the geographic coordinate system through a target binding algorithm, so as to associate and fuse the same target detected in video streams from different cameras. Secondly, this application discloses a highway event detection method based on a multivariate algorithm model, which includes the following steps: Access and collect video stream data from various types of cameras deployed along the highway; The system receives the video stream data and calls multiple pre-set dedicated event detection algorithm models to perform parallel or selective real-time analysis of the video stream data in order to detect traffic events, analyze traffic conditions, and monitor equipment status. The system receives the analysis results from the multi-element algorithm analysis layer and visualizes and alerts the traffic events, traffic conditions, and equipment status.

[0013] Beneficial Effects: The highway event detection system and method based on a multivariate algorithm model proposed in this application integrates multiple video sources through a unified sensing device access layer. Furthermore, through a multivariate algorithm analysis layer that integrates various dedicated event detection algorithm models for pedestrians, parking, congestion, construction, littering, and speeding, it performs parallel or selective real-time collaborative analysis of video stream data. This enables comprehensive, synchronous, and automated detection and assessment of multiple types of traffic events, real-time traffic conditions, and equipment operating status on highways within a single system framework. It overcomes the technical shortcomings of traditional systems, such as limited detection capabilities and isolated functional modules, improving the coverage, accuracy, and overall system efficiency of event detection. This provides unified and powerful technical support for real-time monitoring, proactive early warning, and efficient emergency response on highways. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A structural block diagram of a highway event detection system based on a multivariate algorithm model provided in this application embodiment; Figure 2 A flowchart illustrating the highway event detection method based on a multivariate algorithm model provided in this application embodiment. Detailed Implementation

[0016] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0017] In recent years, to improve the intelligence level of highway network operation and management, the widespread deployment of video surveillance facilities and the use of computer vision technology for automatic event detection have become industry trends. However, in terms of specific engineering practice and technical implementation, existing event detection solutions have inherent architectural flaws that limit their actual effectiveness. Most of these systems are built on isolated, single-function analysis models. For example, some systems can only detect parking or wrong-way driving events through vehicle trajectory analysis, while for completely different anomalies such as road debris, illegal construction, or pedestrian intrusion, they require a separate algorithm or even a separate hardware system for identification. This single-analysis approach leads to fragmented system functionality, forcing highway management to deploy and maintain multiple parallel subsystems to barely cover the main event types. This not only results in high construction and maintenance costs but also creates multiple information silos.

[0018] More importantly, due to the independent and uncoordinated nature of the various detection functions, the system cannot perform unified spatiotemporal correlation and comprehensive analysis of multi-source monitoring videos. When abnormal parking and subsequent slow-moving traffic occur simultaneously on a road, a single parking detection model may only report the parking event without combining it with traffic flow speed data from nearby cameras to intelligently determine the evolving trend of local congestion caused by the parking event. Similarly, for a compliant road construction project, the system may trigger an alarm by recognizing construction vehicles and workers, but it cannot effectively distinguish whether safety facilities such as cones have been placed as required. This deficiency in contextual correlation and comprehensive judgment capabilities means that the output of existing systems is often a one-sided, isolated, and noisy raw signal (such as a large number of duplicate alarms or non-critical alarms), rather than filtered and correlated information that can be used for decision-making. This seriously affects the monitoring center's assessment of the severity of events, the accurate dispatch of emergency resources, and the overall efficiency of operational decision-making, making it difficult to meet the urgent needs of modern highways for real-time, accurate, and integrated intelligent management and control.

[0019] Against this backdrop, this application aims to propose a novel, systematic solution to overcome the aforementioned technical limitations.

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application. Secondly, in this document, the term "comprising" is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0021] Firstly, this embodiment provides a highway event detection system based on a multivariate algorithm model, such as... Figure 1 As shown, the system includes: The sensing device access layer is used to access and collect video stream data from various types of cameras deployed along the highway; The multi-algorithm analysis layer, connected to the sensing device access layer, is used to receive video stream data and call multiple pre-set dedicated event detection algorithm models to perform parallel or selective real-time analysis of the video stream data in order to detect traffic events, analyze traffic conditions, and monitor device status. The integrated application display layer connects to the multi-element algorithm analysis layer to receive the analysis results from the multi-element algorithm analysis layer and to visualize and alarm traffic events, traffic conditions, and equipment status.

[0022] In its implementation, the sensing device access layer is a software service module deployed in a central computer room or cloud platform. It is physically connected to cameras along the highway via a network switch. This layer continuously monitors and receives digital video streams from these cameras. The multi-algorithm analysis layer is a core computing platform, also deployed in the central computer room. It communicates with the sensing device access layer via a high-speed LAN or data center internal network to ensure low-latency acquisition of video stream data. This platform integrates an algorithm scheduling engine and multiple independent algorithm containers, each running a dedicated event detection algorithm model. The scheduling engine can distribute video stream data simultaneously to multiple algorithm models for parallel analysis, or distribute it only to selected models for analysis, based on predefined rules or manual instructions. The integrated application display layer is a web-based application deployed on an application server. It connects to the multi-algorithm analysis layer via an API interface, acquires analysis result data packets in real time, and converts this data into graphs, charts, pop-ups, and sound prompts, displaying them on the monitor screen in the monitoring center.

[0023] Based on the above, this three-layer architecture works collaboratively. The sensing device access layer unifies the input of heterogeneous video sources, the multi-algorithm analysis layer integrates and collaboratively computes multiple detection capabilities under a unified framework, and the comprehensive application display layer provides a unified decision information outlet. This breaks the situation of scattered detection functions and fragmented data flow in traditional systems, enabling comprehensive, synchronous, and automated detection and analysis of multiple types of traffic events, real-time traffic conditions, and equipment operating status within a single system framework, providing a technical foundation for integrated intelligent management and control.

[0024] To address the network bandwidth pressure and analysis latency issues caused by massive video data backhaul centers and improve the system's real-time response capabilities, as an optional implementation method in this embodiment, the multi-algorithm analysis layer includes: A video analytics server cluster, which deploys multiple dedicated event detection algorithm models; Edge computing devices, deployed along highways, are used to preprocess and perform preliminary analysis on video stream data from nearby sensing devices at the access layer, and transmit the analysis results or video feature data to a video analysis server cluster.

[0025] In practice, the video analytics server cluster consists of multiple physical or virtual servers, centrally deployed in provincial or roadside monitoring centers. These servers collectively undertake high-precision, high-complexity analysis tasks, as well as the final fusion and storage of data. Edge computing devices are industrial-grade ruggedized devices integrating computing, storage, and network modules, installed in chassis within communication stations or monitoring poles along highways. Each edge computing device is directly connected to several to dozens of nearby cameras via fiber optic or Ethernet cables. Lightweight analytics software runs on the edge computing devices, capable of preprocessing the received raw video stream, including decoding, frame extraction, and target detection (e.g., using a lightweight YOLO model), extracting structured target metadata (such as location, type, and velocity vectors) or keyframe features. The amount of this metadata or features is much smaller than the raw video stream; the edge computing device encapsulates it and uploads it to the central video analytics server cluster.

[0026] Based on the above, this two-tier architecture of edge preprocessing + central deep analysis reduces the amount of raw data that needs to be transmitted to the center, lowering the demand on backbone network bandwidth. Simultaneously, some simple or highly real-time preliminary analyses (such as moving target detection) can be quickly completed and alerted at the edge, shortening the end-to-end latency of event perception. The central video analytics server cluster utilizes more powerful computing capabilities to fuse, deeply analyze, and assess complex events from metadata from multiple edge nodes, ensuring the overall analytical accuracy and comprehensive decision-making capabilities of the system.

[0027] To meet the extremely high computational resource requirements of multiple dedicated event detection algorithm models running in parallel and in real time, as a further optional implementation method in this embodiment, the video analysis server cluster includes servers using an x86 architecture and configured with NVIDIA GPU accelerator cards or Cambricon smart accelerator cards.

[0028] In the specific implementation, each server in the video analytics server cluster uses a standard x86 architecture processor, such as the Intel Xeon series or AMD EPYC series, to provide general computing power and system compatibility. To accelerate the inference process of deep learning models, each server is equipped with at least one high-performance computing accelerator card. One optional configuration is to use a GPU accelerator card manufactured by NVIDIA, such as the RTX 3090 or A10, which utilizes its CUDA cores and Tensor cores for massively parallel floating-point operations, particularly suitable for convolutional neural network computation. Another optional configuration is to use an intelligent accelerator card manufactured by Cambricon, such as the MLU370-S4, which employs a dedicated instruction set and architecture optimized for AI computing. In practical applications, other models of accelerator cards can also be selected, and this application embodiment does not limit this. The server operating system is Ubuntu 18.04 LTS or higher, with the corresponding accelerator card driver and deep learning inference framework installed.

[0029] Based on the above, this configuration provides powerful heterogeneous computing capabilities for the multi-algorithm analysis layer. GPUs or AI accelerator cards can efficiently perform matrix and tensor operations in deep learning models, enabling multiple dedicated event detection algorithm models to synchronously analyze multiple video streams at high frame rates. This ensures the real-time processing of massive amounts of video data and is a key hardware foundation supporting the efficient operation of the entire system.

[0030] To achieve cross-camera target tracking, accurate location of events, and multi-view data fusion, and to solve the problems of limited field of view and inaccurate positioning of a single camera, as another further optional implementation method in this embodiment, the multi-algorithm analysis layer also includes an online calibration and target spatiotemporal fusion module; The online calibration and target spatiotemporal fusion module is used to calibrate cameras and construct a mapping model between the image coordinate system and the geographic coordinate system through a target binding algorithm, so as to associate and fuse the same target detected in video streams from different cameras.

[0031] In its implementation, the online calibration and target spatiotemporal fusion module comprises a calibration unit and a fusion unit. The calibration unit provides an interactive interface where the operator clicks on at least four feature points with known real-world geographic coordinates (obtained via GPS measurement) in the video frame, such as lane line endpoints or sign corners. The module then uses a direct linear transformation (DLT) or a more complex nonlinear model to solve for the pixel coordinates of the camera image. with geographic coordinates (Assuming the road surface is flat) The mapping relationship between them. For each camera accessing the system, this calibration process must be completed to establish its unique mapping model. The fusion unit receives target detection results from different cameras in real time. Each result includes target image coordinates, type, timestamp, and confidence score. For the detection results of adjacent cameras, the fusion unit first uses their respective mapping models to transform the image coordinates to a unified geographic coordinate system. Then, for targets with similar timestamps, their Euclidean distance in geographic space is calculated. Target type matching degree and similarity of appearance features (e.g., cosine similarity of feature vectors extracted through the ReID model). Using the Hungarian algorithm or a greedy matching algorithm, the optimal matching object is found for each target, with a matching cost function. It can be designed as: ,in This is the weighting coefficient. If the matching cost is lower than the set threshold, it is determined to be the same physical target, assigned a globally unique ID, and its trajectory, speed, and other information are integrated.

[0032] Based on the above, through online calibration and target spatiotemporal fusion, the system can integrate information scattered across multiple camera feeds into a unified geographic information view. This allows the system to track the continuous trajectory of vehicles within camera blind spots, accurately report the specific station number or latitude and longitude of an event, and make more reliable judgments about the same event by integrating information from multiple perspectives. This improves the spatial continuity and positioning accuracy of the system's perception capabilities.

[0033] To achieve comprehensive coverage of various abnormal situations in the complex operating environment of highways, as an optional implementation method in this embodiment, multiple dedicated event detection algorithm models include: pedestrian detection model, parking event detection model, congestion detection model, construction event detection model, non-motorized vehicle detection model, wrong-way driving detection model, litter detection model, vehicle speeding detection model, and vehicle slow speeding detection model.

[0034] In practical implementation, these dedicated event detection algorithms can all employ neural network models trained using deep learning frameworks from existing technologies and deployed in the inference engine of video analytics server clusters or edge computing devices. Pedestrian detection models are used to identify human targets appearing in prohibited areas such as driveways and emergency lanes. Parking event detection models are used to identify vehicles that transition from a moving state to a stationary state and remain stationary. Congestion detection models determine traffic flow interruption status by analyzing vehicle density and average speed within a specific area. Construction event detection models identify specific target combinations related to road maintenance or work sites. Non-motorized vehicle detection models identify two-wheeled vehicles such as bicycles, electric bicycles, and motorcycles that have entered highways. Against-the-way driving detection models determine reverse driving behavior by analyzing the angle between the vehicle's direction of movement and a preset road direction. Spilled object detection models identify static foreign objects such as goods and tire fragments left on the road. Vehicle speeding detection models and vehicle slow-moving detection models calculate instantaneous speed by analyzing the vehicle's displacement between consecutive frames and compare it with a preset threshold.

[0035] Based on the above, by integrating these nine dedicated event detection algorithm models, the system possesses comprehensive detection capabilities covering major types of traffic anomalies. These models are scheduled and coordinated within a unified multi-element algorithm analysis framework, enabling a single system to replace multiple independent and dispersed detection subsystems, thereby expanding the detection scope and integrating system functions.

[0036] To prevent false alarms caused by brief disturbances or normal stops (such as congestion) and to improve the accuracy of parking and pedestrian event alarms, as a further optional implementation method of this embodiment, a pedestrian detection model is used to identify pedestrian targets entering the highway surface within a set detection area, and to generate pedestrian intrusion alarm information when the duration of the pedestrian target's existence reaches a first preset threshold. The parking event detection model is used to identify vehicles that change from a moving state to a stopped state within a highway lane, and generates parking event alarm information when the stopping time of the vehicle exceeds a second preset threshold.

[0037] In practice, the operator can draw polygonal areas on the video screen in the system management interface, designating them as pedestrian detection areas or parking detection areas. The pedestrian detection model uses YOLOv5 or similar algorithms to identify human targets within this area in real time. The system establishes a tracker for each target and times the event; detection only occurs when the target has been present for a certain period of time. Greater than or equal to the first preset threshold (For example, A formal alarm message is generated only after a certain number of seconds.

[0038] The parking event detection model first identifies vehicles using object detection and tracking algorithms, and then uses a Kalman filter to predict and update the vehicle's position in the image. The vehicle at the time instantaneous velocity It can be observed through its image displacement and inter-frame time difference And obtained from calibration parameter estimation. When the system determines Speed ​​below the stop threshold (For example, When the speed reaches (km / h), the timer starts counting down to its stopping time. Only stop time. Exceeding the second preset threshold (For example, A parking event alarm is generated only after a certain number of seconds (in seconds). In practical applications, , and The specific values ​​can be adjusted according to the road section management requirements, and this application embodiment does not limit them.

[0039] Based on the above, the working mechanism of determining duration thresholds effectively filters out instantaneous alarms caused by non-critical situations such as pedestrians briefly crossing the road or vehicles temporarily stopping to yield, reducing the system's false alarm rate. This makes the information reported to monitoring personnel more targeted and reliable, reduces invalid interference, and improves the efficiency of emergency response.

[0040] To improve the rigor of construction event judgment and clarify the lower limit of sensitivity for spill detection, as another further optional implementation method of this embodiment, a construction event detection model is used to simultaneously identify cone barriers, construction vehicles and workers within a set detection area, and generate construction event alarm information when cone barriers, construction vehicles and workers are present at the same time. The spill detection model is used to identify spilled object targets with a size not smaller than a preset pixel size threshold within a set detection area, and to generate a spill event alarm message when the continuous detection time of the spilled object target in the same position reaches a third preset threshold.

[0041] In its implementation, the construction event detection model integrates three parallel target detection sub-networks, used to identify cones (or safety helmets, warning signs), construction vehicles (such as trucks and cranes), and workers, respectively. The system performs detection within designated common construction areas (such as road shoulders and lanes). Let the target sets output by the three sub-networks within one analysis cycle be: cone target set... Engineering vehicle target set Staff target set The model is set with a comprehensive judgment score. The calculation method is as follows: .in, It is an indicator function; its value is 1 when the condition is true, and 0 otherwise. For each type of objective, the weight coefficients are given, and Only when When, that is, all three types of targets are detected simultaneously ( , , Only when the logic determination unit generates a construction event alarm will a construction event alarm be generated.

[0042] The projectile detection model employs a detection network optimized for small targets. The system sets a preset pixel size threshold, such as a minimum area. Pixels. The model only considers the area of ​​the bounding box in the image. The system outputs and tracks suspected projectile targets. It records the target's location. If it is in a continuous Frame (corresponding to a continuous detection time reaching the third preset threshold) ,For example All were detected within seconds, and the location movement distance was [missing information]. Always less than the maximum allowable displacement If the pixel is not clear, it is determined to be a stationary spill and an alarm is generated.

[0043] Based on the above, the construction event detection model effectively distinguishes between compliant construction and simple vehicle malfunctions or unauthorized personnel stops by using joint judgment conditions for the simultaneous existence of multiple targets, thus improving the professionalism and accuracy of alarms. The debris detection model, by setting pixel size thresholds and displacement tolerances, avoids including invalid targets such as worn road markings, shadows, and small pieces of debris moving with the wind in the analysis, focusing instead on stationary obstacles of a certain size or larger that may affect driving safety, thereby controlling computational resources and the number of alarms while ensuring a high detection rate.

[0044] To address the issue of numerous duplicate alarms caused by continuous algorithm detection or events triggered across cameras, and to differentiate the handling of pedestrian events of different natures, thereby improving the effectiveness of alarm information and its decision support value, as a further optional implementation method in this embodiment, the multi-algorithm analysis layer also includes an alarm logic optimization module. The alarm logic optimization module is used to filter or classify alarm information generated by multiple dedicated event detection algorithm models; among them, it suppresses repeated alarm information triggered by the same traffic event, and / or classifies pedestrian intrusion alarm information according to whether the identified person is wearing work clothes with preset characteristics.

[0045] In its implementation, the alarm logic optimization module is a software service running on a video analytics server cluster. It receives the raw alarm information streams generated by all dedicated event detection algorithm models. For the duplicate alarm suppression function, the module maintains a dynamic event pool for each type of event. When a new alarm is received, its key features are extracted, such as event type, spatial location (corresponding to camera ID and image coordinates or geographic coordinates), and occurrence time. The module calculates the feature similarity between this alarm and each active alarm in the event pool. If the similarity exceeds a preset threshold, it is determined to be a duplicate alarm for the same event, and only the duration of the original alarm is updated without generating a new alarm entry. The similarity can be obtained using any calculation method in existing technologies, such as cosine similarity based on feature vectors or weighted calculation based on spatiotemporal distance. For the alarm classification function, when the pedestrian detection model identifies a pedestrian target, a clothing recognition sub-model based on image classification is simultaneously invoked to determine whether the pedestrian is wearing work clothes with highly visible reflective stripes or a specific color. If the system identifies the user as wearing a uniform with preset characteristics, the alarm level will be marked as "Alert Level" in the generated alarm message; otherwise, it will be marked as "Warning Level". Different alarm levels are distinguished by different colors (such as yellow and red) or sounds in the integrated application display layer.

[0046] Based on the above, the alarm logic optimization module purifies the alarm information flow by suppressing duplicate alarms, preventing the monitoring screen from being overwhelmed by repeated information about the same event, and allowing operators to focus on newly occurring events. By classifying pedestrian alarms based on clothing, the system can automatically distinguish between maintenance personnel illegally walking on the road and unauthorized personnel entering the area, providing monitoring personnel with preliminary risk assessment data and supporting them in taking more targeted measures, thus improving the intelligence and practical value of the system's alarms.

[0047] To ensure compatibility with various brands and models of video surveillance equipment already installed on highways and to reduce system access barriers and integration costs, as a further optional implementation method in this embodiment, the sensing device access layer supports access to cameras from different manufacturers via RTSP, ONVIF, or GB28181 protocols; the types of cameras include PTZ cameras, dome cameras, bullet cameras, fisheye cameras, and checkpoint capture integrated machines.

[0048] In its implementation, the sensing device access layer incorporates multiple protocol parsing plugins. For cameras supporting the RTSP protocol, the access layer obtains the video stream URL and pulls the stream by sending standard RTSP commands such as DESCRIBE, SETUP, and PLAY to the camera's IP address and a specified port. For cameras supporting the ONVIF protocol, the access layer first performs device discovery and capability negotiation via Web Services calls, and then obtains the media stream address. For platforms or devices conforming to the GB28181 standard, the access layer acts as a SIP client to register, subscribe to, and receive media streams. This layer's software can automatically or manually adapt to H.264 or H.265 encoding formats. The connected cameras have diverse physical forms: PTZ cameras can be remotely controlled for rotation and zoom; dome cameras typically have a fixed viewing angle; bullet cameras are commonly used for long-distance monitoring; fisheye cameras provide 180-degree or 360-degree panoramic views; and integrated capture cameras combine video surveillance and high-resolution capture functions. In practical applications, there are many choices for the brand, specific model, and encoding format of the camera, and this application does not limit these choices.

[0049] Based on the above, this multi-protocol, multi-type compatible design enables the system to seamlessly integrate existing monitoring resources without the need for mandatory replacement of existing cameras, thus improving the system's applicability and cost-effectiveness. The unified access layer transforms heterogeneous video sources into standardized internal data streams, providing stable and consistent data input for subsequent multi-algorithm analysis.

[0050] This embodiment provides a highway event detection method based on a multivariate algorithm model in its second aspect, which is applied to the highway event detection system based on a multivariate algorithm model as described above, such as... Figure 2 As shown, the method includes the following steps: Access and collect video stream data from various types of cameras deployed along the highway; It receives video stream data and calls multiple pre-set dedicated event detection algorithm models to perform parallel or selective real-time analysis of the video stream data in order to detect traffic events, analyze traffic conditions, and monitor equipment status. It receives the analysis results from the multi-element algorithm analysis layer and visualizes and alerts traffic events, traffic conditions, and equipment status.

[0051] In practice, the method is executed by a software system deployed on a computing device, and the process is as follows: Step S101, Access and collect video stream data: After the system starts, the sensing device access layer initiates connection requests to various cameras deployed along the highway according to the preset device list through RTSP, ONVIF or GB28181 protocol, establishes a stable video stream transmission channel, and continuously collects encoded video stream data.

[0052] Step S102, Multi-algorithm Analysis: The multi-algorithm analysis layer receives the video stream data from step S101. Based on the configuration, the algorithm scheduling engine simultaneously distributes one video stream to multiple dedicated event detection algorithm models, such as the pedestrian detection model, parking event detection model, and congestion detection model (parallel analysis), or distributes it only to a specified model according to task requirements (selective analysis). Each model independently processes the video frames, outputting structured event alarm information or traffic parameters (such as flow rate and speed).

[0053] Step S103, Visualization and Alarms: The integrated application display layer obtains the analysis results generated in step S102 in real time through message queues or API calls. The display layer parses these results, associates them with the GIS map, marks the event locations on the electronic map as icons, displays traffic conditions as color heatmaps, and displays device status in list format. When an alarm occurs, an alarm window automatically pops up, plays a prompt sound, and displays an event screenshot or replay video.

[0054] It should be noted that the highway event detection method based on the multivariate algorithm model in this embodiment corresponds to the aforementioned highway event detection system based on the multivariate algorithm model. Therefore, the parts of the highway event detection method based on the multivariate algorithm model and the storage medium that are not described in detail (including but not limited to specific technical means and effects) can be referred to the relevant descriptions in the aforementioned highway event detection system based on the multivariate algorithm model. This text will not repeat them here.

[0055] In the embodiments provided in this application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the associated hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available medium accessible to a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0056] Finally, it should be noted that the above description is only a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A highway event detection system based on a multivariate algorithm model, characterized in that, The system includes: The sensing device access layer is used to access and collect video stream data from various types of cameras deployed along the highway; The multi-algorithm analysis layer, connected to the sensing device access layer, is used to receive the video stream data and call multiple pre-set dedicated event detection algorithm models to perform parallel or selective real-time analysis of the video stream data in order to detect traffic events, analyze traffic conditions, and monitor device status. The integrated application display layer is connected to the multi-element algorithm analysis layer, and is used to receive the analysis results of the multi-element algorithm analysis layer, and to visualize and alarm the traffic events, traffic conditions and equipment status.

2. The highway event detection system based on a multivariate algorithm model according to claim 1, characterized in that, The multivariate algorithm analysis layer includes: A video analytics server cluster, wherein the video analytics server cluster is equipped with the multiple dedicated event detection algorithm models; Edge computing devices, deployed along highways, are used to preprocess and perform preliminary analysis on video stream data from the access layer of the nearby sensing devices, and transmit the analysis results or video feature data to the video analysis server cluster.

3. The highway event detection system based on a multivariate algorithm model according to claim 2, characterized in that, The video analytics server cluster includes servers that use an x86 architecture and are configured with NVIDIA GPU accelerator cards or Cambricon smart accelerator cards.

4. The highway event detection system based on a multivariate algorithm model according to claim 1, characterized in that, The multiple dedicated event detection algorithm models include: pedestrian detection model, parking event detection model, congestion detection model, construction event detection model, non-motorized vehicle detection model, wrong-way driving detection model, litter detection model, vehicle speeding detection model, and vehicle slow-moving detection model.

5. The highway event detection system based on a multivariate algorithm model according to claim 4, characterized in that, The pedestrian detection model is used to identify pedestrian targets entering the highway surface within a set detection area, and to generate pedestrian intrusion alarm information when the duration of the pedestrian target's existence reaches a first preset threshold. The parking event detection model is used to identify vehicle targets that change from a driving state to a stopped state in a highway lane, and to generate parking event alarm information when the stopping time of the vehicle target exceeds a second preset threshold.

6. The highway event detection system based on a multivariate algorithm model according to claim 4, characterized in that, The construction event detection model is used to simultaneously identify traffic cones, construction vehicles, and workers within a set detection area, and to generate construction event alarm information when all three are present. The spill detection model is used to identify spilled object targets with a size not less than a preset pixel size threshold within a set detection area, and to generate a spill event alarm message when the continuous detection time of the spilled object target at the same position reaches a third preset threshold.

7. The highway event detection system based on a multivariate algorithm model according to claim 1, characterized in that, The sensing device access layer supports access to cameras from different manufacturers via RTSP, ONVIF, or GB28181 protocols; the types of cameras include PTZ cameras, dome cameras, bullet cameras, fisheye cameras, and integrated camera capture and capture devices.

8. The highway event detection system based on a multivariate algorithm model according to claim 4 or 5, characterized in that, The multi-element algorithm analysis layer also includes an alarm logic optimization module; The alarm logic optimization module is used to filter or classify the alarm information generated by the multiple dedicated event detection algorithm models; wherein, it suppresses repeated alarm information triggered by the same traffic event, and / or classifies pedestrian intrusion alarm information according to whether the identified person is wearing work clothes with preset characteristics.

9. The highway event detection system based on a multivariate algorithm model according to claim 2, characterized in that, The multivariate algorithm analysis layer also includes an online calibration and target spatiotemporal fusion module; The online calibration and target spatiotemporal fusion module is used to calibrate the camera and construct a mapping model between the image coordinate system and the geographic coordinate system through a target binding algorithm, so as to associate and fuse the same target detected in video streams from different cameras.

10. A highway event detection method based on a multivariate algorithm model, characterized in that, The method includes the following steps: Access and collect video stream data from various types of cameras deployed along the highway; The system receives the video stream data and calls multiple pre-set dedicated event detection algorithm models to perform parallel or selective real-time analysis of the video stream data in order to detect traffic events, analyze traffic conditions, and monitor equipment status. The system receives the analysis results from the multi-element algorithm analysis layer and visualizes and alerts the traffic events, traffic conditions, and equipment status.

Images