CN-121982663-A - YOLOv13 (13) -based traffic violation detection and linkage reporting method and system

CN121982663ACN 121982663 ACN121982663 ACN 121982663ACN-121982663-A

Abstract

The invention relates to the technical field of intelligent traffic video analysis and law enforcement assistance, and particularly provides a traffic violation detection and linkage reporting method and system based on YOLOv. The method comprises the steps of accessing a video stream, preprocessing, introducing SegNext _attention and SDI feature interaction module to improve a YOLOv network, inputting an image sequence to an improved YOLOv, outputting a target bounding box, a violation category and confidence level, obtaining stability judgment through target association and time sequence verification, triggering a violation event and removing duplication according to a rule base, generating an evidence packet containing a panoramic view and a target cropping map, binding geographical coordinates, uploading, grading, alarming, linking the peripheral equipment, and monitoring the state of equipment for remote configuration. The invention solves the problems of false leak detection, class shake and insufficient engineering closed loop in complex scenes, and improves the traffic violation discovery and treatment efficiency.

Inventors

LIU HAIYING
WANG RUIPENG
JIA MENG
DENG LIXIA
XU ZHENHUA
WU YONGQIANG

Assignees

齐鲁工业大学(山东省科学院)

Dates

Publication Date: 20260505
Application Date: 20260128

Claims (10)

1. The traffic violation detection and linkage reporting method based on YOLOv is characterized by comprising the following steps: s1, video access and preprocessing, namely acquiring images and video streams through video access equipment, performing frame extraction on the video streams, denoising the images and performing scale normalization processing on the images to obtain an image sequence to be detected; s2, improving YOLOv, introducing a SegNext _attention Attention module at a high-level feature output end of a backstone, and introducing an SDI feature interaction module at a multi-scale feature fusion node of Neck to obtain an improved YOLOv network; S3, inputting an image sequence to be detected into an improved YOLOv network, and outputting a bounding box, a violation category and a confidence of a target in each frame of image; S4, performing target association and time sequence consistency verification based on the output result of the step S3 to obtain a stable violation type judgment result; S5, carrying out event judgment and de-duplication combination according to the stable violation type judgment result; S6, when the triggering of the illegal event is judged, evidence preservation and privacy protection are executed, and an evidence packet containing a panoramic image and a target cutting image is generated; s7, performing positioning binding and linkage reporting on the generated evidence packet; s8, setting alarm grades for different violation categories in advance, and carrying out hierarchical alarm by combining the violation categories with confidence, pushing alarm information comprising the violation categories, the occurrence time, the bound position coordinates and the panoramic image and the target cutting image in the evidence package to a law enforcement terminal, and sending alarm signals to an audible and visual alarm or sending control signals to other peripheral equipment; S9, monitoring and reporting the running state of the video access equipment, and carrying out remote configuration and statistical management on parameters and strategies.
2. The traffic violation detection and linkage reporting method based on YOLOv' 13 of claim 1, wherein step S2 includes: S21, the improved YOLOv network is formed by Backbone, neck, head, and a SegNext _attention module is introduced to a high-level feature output end of the backhaul; S22, the SegNext _attention Attention module performs multi-scale space Attention enhancement on the backstone output characteristic, obtains an Attention enhancement characteristic and sends the Attention enhancement characteristic to Neck; S23, introducing an SDI feature interaction module into a multiscale feature fusion node of Neck; S24, the SDI feature interaction module performs scale alignment on different scale features participating in fusion, and performs element-by-element interaction fusion after convolution transformation to generate fusion features.
3. The traffic violation detection and linkage reporting method based on YOLOv' 13 of claim 1, wherein in step S3, the violation categories of the targets in each frame of image include: Red light running violations, no wearing of safety helmets, violations of traffic or passenger traffic, violations of pedestrian traffic, violations of crosswalk, violations of crossing roads, and normal traffic.
4. The traffic violation detection and linkage reporting method based on YOLOv as claimed in claim 1, wherein in step S4: And (3) executing target association on the output result of the step (S3), establishing a cross-frame corresponding relation of the same target through a Hungary algorithm or an IOU matching algorithm, setting a fixed time window, and adopting majority voting or sliding average smoothing processing on the class confidence coefficient of the same target in the window to obtain a stable violation class judgment result.
5. The traffic violation detection and linkage reporting method based on YOLOv as claimed in claim 1, wherein step S5 includes: s51, a rule base containing a trigger threshold, duration of an offending event and duration frame number parameters is constructed in advance; S52, triggering and judging the rule base to judge the rule cases of the stable rule violation type according to the rule base, if the same target repeatedly triggers the same rule cases within the set time, executing de-duplication processing, and if the same rule cases are continuously triggered, merging the event cases into an event segment comprising the starting time and the ending time.
6. The traffic violation detection and linkage reporting method based on YOLOv as claimed in claim 1, wherein the step S6 includes: S61, when the triggering illegal events are judged, automatically capturing the current frame or capturing key pictures in an event section; S62, cutting out a target area based on a target boundary box, and generating an evidence packet containing a panoramic image and a target cutting image; s63, embedding a time stamp, a device number, violation categories and confidence information into the evidence packet, and if privacy protection is required, executing blurring or shielding processing on the face and license plate privacy area in the evidence packet.
7. The traffic violation detection and linkage reporting method based on YOLOv as claimed in claim 1, wherein the step S7 includes: s71, pre-inputting geographic coordinates and optional direction information of monitoring points or acquisition equipment; s72, carrying out association binding on the information and the evidence packet, and uploading the bound evidence packet to a platform server through a network; S73, detecting a network connection state in real time, if network abnormality is detected, storing the evidence packet into a local cache queue, recording an uploading state, and automatically supplementing and transmitting according to a cache sequence after the network is recovered.
8. The traffic violation detection and linkage reporting method based on YOLOv as claimed in claim 5, wherein step S8 includes: s81, presetting alarm priorities for different violation categories; S82, determining a final alarm level by combining the violation category, the confidence level and the duration of the violation event; S83, pushing alarm information comprising violation categories, occurrence time, bound position coordinates and evidence pictures to a law enforcement terminal, and sending control signals to the audible and visual alarm and the traffic signal lamp controller appointed peripheral if the peripheral needs to be linked.
9. The traffic violation detection and linkage reporting method based on YOLOv' 13 of claim 1, wherein step S9 includes: s91, periodically monitoring the offline, frame rate abnormal, picture freezing, overexposure or overdrawing and shielding states of the camera, and reporting the monitoring result in real time; S92, receiving a remote parameter configuration instruction, updating the frame extraction rate, the detection threshold value and the time window strategy parameters, and immediately taking effect and saving configuration after updating.
10. A traffic violation detection and linkage reporting system based on improvement YOLOv for implementing the method of any of claims 1-9, the system comprising: The video access and preprocessing module is used for video access and preprocessing, acquiring images and video streams through video access equipment, extracting frames from the video streams, denoising the images and performing scale normalization processing to obtain an image sequence to be detected; The model improvement module is used for improving YOLOv, introducing a SegNext _attention Attention module at the high-level feature output end of the backstene, and introducing an SDI feature interaction module at the multi-scale feature fusion node of Neck to obtain an improved YOLOv network; the target detection module is used for inputting the image sequence to be detected into the improved YOLOv network and outputting the bounding box, the violation category and the confidence of the target in each frame of image; the target association and time sequence consistency verification module is used for carrying out target association and time sequence consistency verification based on the output result of the target detection module to obtain a stable violation type judgment result; The event judgment and de-coincidence module is used for carrying out event judgment and de-coincidence combination according to the stable violation type judgment result; the evidence collection and preservation and privacy processing module is used for executing evidence preservation and privacy protection when the triggering of the illegal event is judged, and generating an evidence packet containing a panoramic image and a target cropping image; the positioning binding and linkage reporting module is used for executing positioning binding and linkage reporting on the generated evidence packet; The hierarchical alarm and peripheral linkage module is used for presetting alarm grades for different violation categories, combining the violation categories with confidence level, carrying out hierarchical alarm, pushing alarm information comprising the violation categories, the occurrence time, the bound position coordinates and the panoramic image and the target cutting image in the evidence package to the law enforcement terminal, and sending alarm signals to the audible and visual alarm or sending control signals to other peripheral equipment; And the equipment health monitoring and remote configuration module is used for monitoring and reporting the running state of the video access equipment and carrying out remote configuration and statistical management on parameters and strategies.

Description

YOLOv13 (13) -based traffic violation detection and linkage reporting method and system Technical Field The invention relates to the technical field of intelligent traffic video analysis and law enforcement assistance, in particular to a traffic violation detection and linkage reporting method and system based on YOLOv. Background In the conventional road traffic management, traffic offences related to pedestrians and non-motor vehicles mainly depend on manual inspection or post-sampling inspection, and the problems of untimely discovery, limited coverage range, high evidence obtaining cost and the like exist. Along with the popularization of road monitoring equipment, automatic detection based on computer vision becomes an important direction, but the prior art is often influenced by factors such as complex background, shielding, scale change, uneven illumination at night, video jitter and the like in an actual road scene, so that false detection, omission and class jitter are easy to occur, meanwhile, a plurality of schemes only stay in detection output, lack of complete detection, identification and notification capability, and are difficult to meet engineering deployment and law enforcement evidence obtaining requirements. YOLOv13 has better real-time detection capability, but has the defects that the model is influenced by complex background textures, night dazzling light and shielding, the space attention of the model to a key area is insufficient, false detection and confidence fluctuation are easy to occur, meanwhile, targets such as pedestrians, helmets and the like are often long-distance small targets, and cross-scale semantic interaction is insufficient during multi-scale fusion, so that the small targets are easy to miss detection or unstable category. Therefore, an integrated scheme is needed that can realize stable detection in complex road scenes and has the capabilities of evidence solidification, reporting linkage, off-grid repair transmission, equipment health monitoring and the like. Disclosure of Invention In view of the above, in order to solve the problems of false detection, omission, category shake, insufficient engineering closed-loop capability and the like existing in automatic identification of pedestrian and non-motor vehicle traffic violations in a road monitoring scene in the prior art, the invention provides a YOLOv-based traffic violation detection and linkage reporting method and system, which are used for carrying out real-time detection, evidence solidification, positioning binding and alarm linkage on traffic participant behaviors in a road monitoring image or video, introducing SegNext _attribute to a high-level output end of a backbond of YOLOv for carrying out multi-scale spatial Attention enhancement, introducing SDI to realize scale alignment and feature interaction fusion at Neck fusion nodes, and improving detection reliability of a small target and a shielding scene. In a first aspect, the present invention provides a traffic violation detection and linkage reporting method based on YOLOv, where the method includes: s1, video access and preprocessing, namely acquiring images and video streams through video access equipment, performing frame extraction on the video streams, denoising the images and performing scale normalization processing on the images to obtain an image sequence to be detected; S2, improving YOLOv, introducing a SegNext _attention Attention module at a high-level feature output end of a backstone, and introducing an SDI feature interaction module at a multi-scale feature fusion node of Neck to obtain an improved YOLOv network; S3, inputting an image sequence to be detected into an improved YOLOv network, and outputting a bounding box, a violation category and a confidence of a target in each frame of image; S4, performing target association and time sequence consistency verification based on the output result of the step S3 to obtain a stable violation type judgment result; S5, carrying out event judgment and de-duplication combination according to the stable violation type judgment result; S6, when the triggering of the illegal event is judged, evidence preservation and privacy protection are executed, and an evidence packet containing a panoramic image and a target cutting image is generated; s7, performing positioning binding and linkage reporting on the generated evidence packet; s8, setting alarm grades for different violation categories in advance, and carrying out hierarchical alarm by combining the violation categories with confidence, pushing alarm information comprising the violation categories, the occurrence time, the bound position coordinates and the panoramic image and the target cutting image in the evidence package to a law enforcement terminal, and sending alarm signals to an audible and visual alarm or sending control signals to other peripheral equipment; S9, monitoring and reporting the running state of the video access equ