| Detecting abnormal events in the field of transportation plays an important role in preventing and timely handling traffic accidents,and improving the level of urban governance.However,existing methods for detecting traffic abnormal events have low accuracy,poor realtime performance,and limited coverage.To address these issues,this paper proposes a novel vehicle-mounted method for detecting traffic abnormal events.The method uses video information from cameras mounted on intelligent vehicles to perform real-time detection of events such as crowd gathering,vehicle gathering,traffic accidents,and fires in a full-scene context.In the traffic abnormality detection algorithm,the efficiency of the model is improved by adding an Atrous Spatial Pyramid Pooling module,introducing a fusion attention mechanism,and improving the bottom-up path aggregation network.In the target tracking algorithm,the adaptive noise matrix is improved,and a new overlap ratio metric is used to improve the accuracy of vehicle and pedestrian target tracking.Experimental results show that the proposed method has significantly improved detection accuracy and speed.Compared with YOLOv3 and YOLOv5 s models,the m AP value relative to YOLOv3 and YOLOv5 s has increased by 9.2%and 4.3%,respectively.The improved Deep SORT model has increased the MOTA and MOTP by 4.9% and 1.4%,respectively.By using TensorRT for acceleration deployment,the inference time on the mobile end has been reduced by 10 milliseconds,and the FPS value has reached 29,which meets the requirements of real-time detection.In summary,this paper proposes a new solution to the problems of existing traffic abnormal event detection methods,which significantly improves detection accuracy and realtime performance.The research results provide a convenient method for urban managers to effectively warn and respond to traffic abnormal events,improve the level of intelligent transportation management,and provide safer guarantees for urban residents. |
PDF Full Text Request