A Vehicle Counting Method For Traffic Surveillance Video Of Complex Scene Based On Deep Learning

Urban traffic flow statistics is an important part in road traffic state control and management,and it is the basis of smart traffic analysis.As the scale of cities continues to expand,traffic roads present a variety of complex forms,how to accurately monitor video in a large number of congested and complex scenes Traffic flow statistics has become one of the research hotspots in the field of smart transportation.Traditional vehicle flow statistics algorithms such as background difference,frame difference and target detection are not effective when faced with complex scenes of densely congested road traffic.To this end,this paper uses a method based on multi-target tracking and vehicle distribution density estimation to design a smart traffic flow detection system.Aiming at the problem of multi-lane segmentation in complex scenes,an unsupervised multi-lane segmentation method is designed;for the problem of vehicle detection accuracy in complex congestion scenes,an auxiliary vehicle counting method based on vehicle distribution density heat maps is proposed,through multitarget tracking Combining the vehicle distribution density heat map to detect and estimate the traffic flow in the scene,the main tasks are as follows:(1)Unsupervised multi-lane segmentation.Aiming at multiple types of complex road scenes such as intersections,forks,and multiple lanes,an unsupervised multi-lane segmentation method is proposed.The experimental results show that in the face of various complex roads and traffic congestion scenes,it is possible to distinguish lane areas in different directions(2)Target detection.The research has learned the current commonly used multi-target detection algorithms,and used the collected vehicle target detection data sets to perform migration training on the network model,combined with the Non-Local idea,and added relevant modules to improve the detection performance of the target detection algorithm.Experimental results show that the improved Non-Local-based vehicle detection network can obtain an AP score of 93.1 vehicle detection accuracy at the expense of lower inference speed,which is 3.21% higher than the original detection network.(3)A statistical method of vehicle flow combined with a heat map of vehicle distribution density.Aiming at the problem of decreased statistical accuracy caused by severe occlusion in multi-lane congestion,incomplete target and other reasons,the accuracy of vehicle detection is decreased,and an auxiliary statistical method based on the heat map of road vehicle distribution density is proposed.The vehicle density heat map is obtained by adding the density heat map extraction branch to the original target detection network,and the vehicle density heat map is used to obtain the number of vehicles in the scene to assist in traffic flow statistics.