Research And Implementation Of Multi-vehicle Real-time Tracking Method In Traffic Monitoring Scenario

At present,China's urbanization is developing at a high speed.With the rapid growth of the number of motor vehicles,problems such as traffic congestion,traffic management,and accident rescue in urban development have become increasingly prominent.With the rapid development of big data and deep learning,smart transportation and smart city are also the inevitable trend of current urban development.Smart transportation is an indispensable part of smart city.Vehicle tracking is the basis of traffic congestion prediction and traffic intelligent regulation.At present,there are relatively few related studies on real-time tracking of multiple vehicles in traffic monitoring scenarios.How to track multiple vehicles in real time and accurately has always been a problem.This paper studies the multi-vehicle tracking method in the traffic monitoring scenario.It will study the real-time multi-vehicle tracking from three aspects:vehicle detection,vehicle data association,and single vehicle tracking.Firstly,we collected the existing vehicle data set in traffic monitoring scenario,and on this basis,we added a newly collected data set to expand the vehicle detection data set.For the real-time detection of vehicles,this paper proposes a vehicle detection algorithm based on key points and integrates the self-attention mechanism module into the detection algorithm to improve its accuracy.Secondly,combined with the characteristics of the traffic monitoring scene,this paper propose vehicle association algorithm with multi-dimensional information fusion.The results of vehicle detection are used as the input of vehicle re-identification to extract the apparent characteristics of the vehicle,while combining the historical trajectory and spatial position information of the vehicle.Making the vehicle's frame-to-frame association more accurate.For missed vehicles,this paper uses the single-target tracking method based on the twin network to predict,and uses asymmetric convolution and depth separable convolution to trim the tracking model,which improves the single-vehicle tracking's speed with less loss of accuracy.Finally,a multi-vehicle tracking prototype system was designed and developed,integrating the vehicle detection,vehicle association and single vehicle tracking method proposed above.It can visualize vehicle tracking results,output vehicle driving directions and traffic statistics,and can record trajectory information.