Multiple Object Tracking In Complex Traffic Videos Based On Deep Learning

With the rapid growth of vehicle ownership and the continuous improvement of traffic monitoring in China,how to utilize these videos with computer vision technology to obtain traffic information has become a concern in the intelligent traffic system.The key technology of acquiring traffic information is vehicle tracking.With the popularization of GPUs and the success of deep learning in object detection task without manually setting parameters,the method of deep learning is introduced into multi-object tracking(MOT).Therefore it has very important practical significance for the continuous and stable object tracking of complex traffic videos in large cities.In this paper,we incorporate appearance features and motion features extracted from the deep neural network into the data association section of MOT algorithm.Additionally,we improved the MOT algorithm for the missed detection problem of short-term detection algorithms.The main work of the paper is as follows:(1)In this paper,a convolution neural network model is designed to extract the distinctive appearance features while considering the demand of real-time tracking and the characteristics of similar appearance.By training Siamese networks to solve the challenge of the small difference in the class and large intra-class difference in complex traffic scenes.In addition,different from other convolutional neural network-based appearance feature extraction methods,we reduce the appearance features to two-dimensions by the final full-connected layer.Thereby a more compact descriptor is obtained on the premise of ensuring the discrimination.The experimental results on the public dataset show that our method achieves the tracking accuracy of 71.9%and reduces the false match rate of the target by 11%.(2)Aiming at the existence of short-time occlusion situation and the nonlinear motion pattern in complex traffic videos,we design a highly efficient GRU network to extract the motion feature.According to the tracking results of the above tracking algorithm,we manually select suitable training set and train GRU network for speed estimation.Since the velocity prediction model is updated online for each target,it can capture the specific nonlinear motion pattern of the specific target more accurately,so the tracking accuracy is significantly improved.Moreover,this model is applied to occlusion and missed detection to further improve the tracking performance.The experimental results on the public dataset show that the MOT algorithm with two features is better than the algorithm only adding the appearance feature,and the target mismatches are further reduced by 26%.We conducted extensive experiments on UA-DETRAC traffic video dataset that collected from Beijing,Tianjin,and other places in China.The model is time-consuming when extracting appearance features,but overall it satisfies the real-time demands.Compared with the two top MOT algorithms in the benchmark,our algorithm achieves the highest tracking accuracy,and the number of mismatches while tracking is far less than the other methods.