Research On Multi-target Tracking Technology Based On Recurrent Neural Network

Multi-target tracking technology is mainly used in the military and civil fields.Tracking multiple targets in an unconstrained environment faces multiple challenges,including a priori unknown and time-varying number of targets,a continuous state estimation of all targets,and a discrete data association issue.Recently,with the rise of deep learning,deep neural networks have achieved good results in vision-based tracking.However,deep learning is mostly used for single-target tracking or vision-based multi-target tracking,and there are not many methods in multi-target tracking in non-visual scenes.This thesis studies how to use recurrent neural network(RNN)to solve several key problems in multi-target tracking.The main research work of this thesis is as follows:(1)A target state prediction algorithm based on Long Short-Term Memory(LSTM)is proposed.An LSTM-based network is used at this algorithm to learn the timing rules of a large number of individual target motion data to achieve the prediction of a single target state.Experiments show that the algorithm can effectively predict the target motion state,and its accuracy is higher than traditional filtering,and has better adaptability to maneuvering targets.(2)Aiming at the difficulty in multi-target tracking-data association problem,a data association algorithm based on LSTM is proposed.An LSTM-based model can completely learn from the data and output the assignment probability of the target and the measurement,which also consider the situation that the there isn't any measurement associated with the target.Good results have been obtained under the conditions of different clutter density and detection probability.The algorithm takes less time and has higher correlation accuracy.(3)Design and implement a multi-target tracking method based on RNN and JPDA.At each time step,the method divides the multi-target tracking problem into three stages,data association,target state prediction,and track management.Where the track management module adopts an RNN-based track management method proposed in this thesis.The simulation results show that the proposed method can solve the problem of multi-target tracking with clutter and time-varying target number effectively.