Multi-object Tracking Algorithm Based On Graph Convolution Network

Online multi-object tracking technology plays an important role in real-time video scene analysis,such as machine navigation,autonomous driving and video monitoring system.As the GPU computing capacity of ascension and deep learning network application in object detection,the tracking by detection has been becoming the current main research methods,most of the multi-object tracking algorithm was conducted on the basis of testing data correlation,therefore,it is very challenging to correlate the object detection results of complex scene video with the existing trajectory.Currently,the commonly used Hungarian algorithm and KM matching algorithm have a strong dependence on detection algorithm.In the scene with complex,changeable motion and frequent occlusion,the number of identity conversion is high,and the phenomenon of trajectory fragmentation is easy to occur,unable to cope with various practical scenarios.In this paper,the convolutional neural network is used to extract the appearance features of the object,and the graph convolution neural network is used to solve the bipartite graph matching.In order to directly improve the evaluation indexes MOTA and MOTP of multi-objective tracking,we also use the guidable Hungarian algorithm to realize the end-to-end multi-object tracking algorithm.The main work of this paper is as follows:(1)Aiming at target in complex scenes overlap block phenomenon,this paper designed to extract the distinction between the detected object appearance characteristics of convolution neural network model,build objects match the figure structure,use graph convolution network to solve the weighted graph matching data association,according to the real value of matrix and graph matching network and output hypothesis estimation matrix multi-stage matrix loss function,and MOT17 data set was used for parameter optimization;(2)to improve the evaluation indexes MOTA and MOTP of multi-object tracking,a guidable Hungarian algorithm based on deep recurrent neural network was proposed to realize the end-to-end deep neural network multi-object tracking algorithm.The traditional Hungarian algorithm is not derivable and cannot be used to carry out back propagation by neural network.Therefore,a new matching algorithm is defined in this paper,which simulates the Hungarian algorithm by two-way recurrent neural network and uses the derivable loss function to carry out network back propagation.Proposed two algorithms in MOT17 and MOT15 test data sets,the experimental results show that the proposed two algorithms were superior to the existing four kinds of algorithms,among them,the multiple target tracking algorithm based on figure convolution on MOT17 dataset ID?Sw identity exchange frequency decreased obviously,compared with the online RNN?LSTM end-to-end multiple target tracking algorithm,ID?Sw increased by 26.6%,compared with the algorithm based on differentiable Hungary algorithm on the MOTA,and MOTP index increased by 5.1% and 4.4%respectively,and the tracking effect for specific video display.