Research On The Algorithm Of Urban Road Object Detection And Tracking Based On Deep Learning

With the improvement of people’s living standards and the enrichment of social activities,the number of vehicles on urban roads has also increased.While it brings convenience to people,it also brings problems such as traffic congestion and frequent traffic accidents.In order to avoid the above-mentioned problems as much as possible,the urban road intelligent monitoring system is introduced in urban road traffic management.It can detect vehicles or pedestrians appearing on the road from real-time surveillance video,and quickly convey the relevant information of vehicles or pedestrians with abnormal behaviors on urban roads to relevant personnel for processing.As urban roads are located in the outdoor scene,there are usually more vehicles and pedestrians on the road,so it is easy to have problems such as target occlusion and small target.The problems mentioned above are likely to cause missed inspection of the target.Therefore,based on the theory of deep learning,this paper uses the urban road target detection algorithm based on YOLO V3 to detect vehicles or pedestrians in urban roads on the premise of ensuring the real-time performance of urban road target detection algorithm and aiming at improving the detection rate and recognition rate of occluded objects and small targets.A processing method based on Soft-NMS filter box is proposed to improve the detection rate of occluded objects.Compared with the original YOLO V3 on the self-made data set,m AP is improved by 2.3%.And there is also a 1.1% increase in public data sets.At the same time,the target tracking algorithm with Deep-SORT as the core is introduced.When Deep-Sort algorithm is used for target tracking,the tracked target is easy to be interfered by similar targets around,and the tracked target fails to be tracked when its appearance changes greatly.In view of this situation,on the basis of the overall framework of Deep-Sort target tracking algorithm,a multi-position prediction matching method based on Siamese Network of urban road target similarity is proposed to improve the accuracy of matching between trajectory and prediction box in the tracking process,and finally achieve the purpose of improving the tracking effect.The experimental results show that the algorithm adopted in this paper improves the detection rate and tracking rate of obstructed targets of urban roads while ensuring real-time performance.