| In the intelligent transportation system,vehicle detection and tracking based on road monitoring is of great significance for improving the comprehensive management level of urban traffic,identifying abnormal vehicle behavior,and implementing traffic violation control.However,the task of vehicle detection and tracking is affected by factors such as complex urban traffic scenes,vehicle occlusion,and multi-scale changes of vehicles,and there are problems such as false detection,missed detection,and tracking loss.In view of the above problems,this paper improves the vehicle detection and tracking algorithm on urban roads based on road monitoring.The main research results are as follows:(1)A vehicle detection algorithm based on improved YOLOv5 s is proposed.Firstly,by reconstructing the multi-scale fusion network and the output end of the head network,the utilization rate of shallow features of the backbone network is improved,and the detection ability of the algorithm for targets of different scales is enhanced.Secondly,the CBAM attention mechanism is embedded before each feature information fusion in the Neck network,which enhances the algorithm’s attention to the deep semantic information of the feature map,and improves the vehicle detection ability of the detection algorithm in the scenarios of occlusion and environmental interference.Finally,the SIo U loss function considering the regression direction is introduced to reduce the degree of freedom of the detection frame regression,and improve the convergence speed and detection accuracy of the algorithm.(2)An improved Deep SORT vehicle tracking algorithm based on it is proposed.Aiming at the problem of insufficient appearance feature extraction ability of Deep SORT algorithm,a Deep SORT vehicle tracking algorithm based on Res Net18 is proposed,that is,the residual network Res Net18 is used as the appearance feature extraction network of the tracking algorithm,which improves the extraction ability of the tracking algorithm on the appearance features of the vehicle by deepening the network level,replaces the vehicle detector in the Deep SORT algorithm with the improved YOLOv5 s algorithm,and adjusts the tracking parameters according to the urban road scene.Finally,the Veri-wild dataset is used to train the improved appearance feature extraction network,and the effectiveness of the vehicle tracking algorithm is verified in different road scenarios.(3)A lane change trajectory prediction method based on vehicle detection and tracking model is proposed.The R-Deep SORT algorithm is used to extract the tracking trajectory of the vehicle,and the interaction mode of Open CV is used to extract the lane line information in the monitoring area.The lane change model adopts the lane change trajectory prediction algorithm of LSTM based on the attention mechanism,that is,the attention mechanism is introduced to increase the weight of the LSTM time series network on the historical key trajectory indicators,and finally the effectiveness of the method is verified on the urban road monitoring data collected in the field.In summary,this paper proposes a vehicle detection and tracking model for urban roads and applies research on the prediction of vehicle lane change trajectory,and experiments show that the proposed improved algorithm meets the needs of vehicle detection and tracking under urban road monitoring,and has certain reference value for the intelligent management of urban traffic. |
