Research On Video Moving Vehicle Detection Based On Deep Learning

With the rapid development of economy and the increasingly complex traffic environment,the demand of establishing intelligent transportation has become considerably urgent.The object detection technology based on deep learning has been widely applied to monitor the vehicles in the traffic,but the object detection on occlusion can't make full use of the time sequence information in the video stream,and the detection accuracy of moving objects in the video is low.In order to solve this problem and make full use of the advantages of the current object detection network,this paper proposes and optimizes the LSTM-YOLOv3 object detection algorithm which integrates LSTMs,and conducts experimental verification on the surveillance video data set UA-DETRAC.The primary contributions of this paper are listed as follows:Firstly,a LSTM-CONV framework with good scalability has been put forward.Based on the framework,a LSTM-YOLOV3 fusion object detection algorithm has been proposed.The algorithm takes into account both the multi-scale detection effect of YOLOv3 and the temporal and spatial continuity of LSTM,making it more suitable for the application on video stream object.The experimental results show that the accuracy of the proposed fusion network model is improved by 5%,and the effect of detection in the case of occlusion and blur is pretty well.Secondly,the LSTM-YOLOV3 fusion algorithm proposed in this paper has been ameliorated.In this paper,some new strategies such as mosaic data enhancement,K-means++clustering algorithm,CIOU Loss,Focus Loss,Si Lu activation function are incorporated into the algorithm to improve the network performance.Mosaic data enhancement strategy plays a role in the improvement of the detection accuracy of the network;K-means + + clustering algorithm is used to generate anchor frame more suitable for vehicle detection;CIOU Loss and focus loss are used to speed up the convergence speed of the network and solve the problem of positive and negative sample imbalance respectively;Si Lu activation function enhances the nonlinear expression ability of the network model.Experimental results show that the improved fusion algorithm is effective,and the m AP of the proposed and optimized fusion detection algorithm is 91.2%.