Research On Vehicle Collision Traffic Accident Detection Based On Deep Learning

Road traffic accidents are a social issue,and their harm affects various aspects of social life.Timely detection and elimination of traffic accidents have always been important research topics in traffic safety management.However,due to the complexity of the traffic environment and the limitations of traditional detection methods,the actual effectiveness of vehicle collision traffic accident detection could often be better.With the rapid development of deep learning technology,the method of vehicle collision detection based on deep learning with drones as the carrier,equipped with imaging equipment and image transmission modules,has become a research hotspot in traffic accident detection.Compared with traditional traffic accident detection methods,deep learning-based detection methods are faster,more accurate,and more robust.Therefore,this paper studies vehicle collision traffic accident detection based on deep learning technology and drone video data.The main research contents are:First,the occurrence process of vehicle collision traffic accidents and the current mainstream target detection and collision detection methods are described and sorted out: the requirements of vehicle collision detection algorithms and the research focus of vehicle collision detection algorithms are clarified;the vehicle target detection model based on the deep learning algorithm YOLOv5 and the OBB collision detection algorithm is used for vehicle collision detection.Second,the construction of the vehicle target detection dataset and the vehicle collision detection dataset is carried out: in response to the problem of fewer open-source datasets under the drone perspective and the lack of vehicle collision traffic accident datasets,the UAV-ROD dataset and UCAS-AOD dataset are used as the basis,combined with a web crawler,data enhancement,and other technologies,to establish a vehicle target detection dataset My car-dataset that meets the experimental requirements of the paper;based on the Beam NG.drive collision simulator,simulation experiments are designed,simulating various collision scenarios,and generating vehicle collision detection datasets.This prepares for the experimental verification of the model designed in the paper.Then,a vehicle collision traffic accident detection model is designed based on the YOLOv5 target detection algorithm: to address the boundary issues brought about by bounding box regression,the YOLOv5 algorithm is improved based on CSL circular encoding technology,realizing vehicle target detection under a top-down perspective.The optimal model configuration is determined by analyzing the impact of the window function radius and the backbone network on model performance.The proposed vehicle target detection model achieves an m AP of 97.9%and an FPS of 57.80 f/s on the self-built dataset—My car-dataset.Comparative analysis with similar target detection models shows that the performance advantage of the designed vehicle target detection model is relatively significant;to address the problem of low detection accuracy for small targets,the CBAM attention mechanism is added and combined with the C3 module to optimize the vehicle target detection model,improving the detection effect of the model for small targets;on this basis,a vehicle collision traffic accident detection model is designed based on OBB bounding box theory;and experimental verification is carried out using the simulated vehicle collision dataset,showing that the designed vehicle collision detection model can effectively achieve collision detection under different traffic scenarios.Finally,based on the vehicle collision detection model designed in the paper,the Py Qt5 is used to design a GUI graphical interface and build a vehicle collision traffic accident detection system.The built vehicle collision traffic accident detection system is tested.The test results demonstrate that the system can effectively accomplish real-time detection and alarming of vehicle collision traffic accidents,meeting the anticipated outcomes..This paper is guided by serving practical work and empowers road traffic safety management with technology.The system can be installed and used directly after packaging,and the research results serve the actual work of public security traffic management.