Research On Traffic Vehicle Tracking Algorithm Based On Multi-domain Convolutional Neural Network

The Intelligent Transportation System can integrate various traffic information on the road accurately and efficiently,improve transportation efficiency,and play an important role in road traffic.Traffic vehicle tracking is an important part of the Intelligent Transportation System.By tracking the target vehicle in the traffic video,information such as the position,running speed,running acceleration,running direction,etc.of the vehicle can be obtained,and then the vehicle's red light,illegal parking,speeding and other irregularities are analyzed.Therefore,it's particularly important for the development of Intelligent Transportation Systems to track vehicles accurately and steadily.However,the actual traffic scene is usually complicated.The vehicles in the video are often interfered with by a series of factors such as similar target vehicles and scale changes during driving,which affect the accuracy and robustness of vehicle tracking.Based on this,this paper launched a study on traffic vehicle tracking.The main research contents and results of the paper are as follows:The Multi-Domain Convolutional Neural Network is introduced into traffic vehicle tracking.However,because it can not effectively deal with similar vehicle interference and scale changes,an improved multi-domain convolutional neural network vehicle tracking algorithm based on DenseNet and attention mechanism is proposed.First,replacing the backbone network of the original Multi-Domain Convolutional Neural Network with two dense blocks and a translation layerof the DenseNet121 network front end,and the features of different layers can be merged by using dense connections to extract more subtle features and richer spatial information to achieve precise positioning of the target vehicle and improve the impact of dimensional changes on the tracking effect.Then,the channel attention mechanism is added to the network,which is advantageous to extract more critical target vehicle information by feature recalibration,and the resolution of the model to the target vehicle is increased.Finally,the vehicle tracking algorithm of Multi-Domain Convolutional Neural Network and the improved Multi-Domain Convolutional Neural Network vehicle tracking algorithm are validated by using part of the vehicle videos in the OTB100 data set,and the experimental results of the two algorithms are compared and analyzed.Because in the actual Intelligent Transportation System,the most commonly usedmonitoring means is the fixed-point surveillance cameras fixed beside the road.So,it is very important to verify the tracking effect of vehicle tracking algorithm in vehicle video captured by fixed-point surveillance camera.In this paper,three vehicle videos are selected to test the Multi-Domain Convolutional Neural Network vehicle tracking algorithm and the improved Multi-Domain Convolutional Neural Network vehicle tracking algorithm,which are collected under the actual road conditions.And the experimental results are compared and analyzed.