Research On Reinforcement Learning Optimization Method Of GLMB Filter Parameters For Traffic Monitoring Scenarios

With the development of intelligent transportation systems and autonomous driving technologies,multi-target tracking algorithms such as Generalized Labeled Multi-Bernoulli Filter(GLMB)have become one of the key technologies in the above two fields.In actual traffic scenarios,millimeter-wave radar multi-target tracking technology faces many challenges,such as changes in scene density,time-varying number of targets,and occlusion between targets.This paper mainly aims at the problem that the GLMB filter parameters cannot match the scene in the time-varying process of the millimeter-wave radar traffic monitoring scene,and it is difficult to ensure the multi-target tracking performance of the filter,including the low accuracy of target tracking and the inaccurate estimation of the number of targets.And the correct rate of track label association is low.The research in this paper shows that among many filter parameters,the detection probability,the newborn probability,and the survival probability are more sensitive to changes in the density of the observed objects,and have a greater impact on the track quality,and need to be optimized and adjusted independently according to the scene changes.To solve this problem,this paper introduces the Long Short Term Memory network(LSTM)and reinforcement learning method into the field of GLMB filter parameter estimation based on typical traffic scenarios,the characteristics of millimeter-wave traffic monitoring radar data and the principle of GLMB filter.To solve the problem of autonomous estimation of filter parameters in changing scenarios,the main achievements are as follows:(1)Based on the raw data of millimeter-wave radar actually collected in three typical scenarios,the theoretical analysis of the parameter combination of GLMB filter is carried out.When detection probability parameters,newborn probability parameters and survival probability parameters do not match with the traffic scene,the influence mechanism on target tracking accuracy,target number estimation,track label association and typical characteristics of false tracks are summarized.(2)In response to the requirement of self-optimization of filter parameters as the scene changes,this paper regards the parameter estimation problem as a sequence decision problem,and proposes a method to infer the filter parameters based on the input millimeter wave radar original point trace data by using the LSTM neural network.Using the actual collected multi-scene radar data for analysis and verification,the above method does not require filters to participate in training,and can independently correct the filter parameter combination value to reflect the scene change trend in real time,but the accuracy is still not accurate enough,and it is difficult to fully meet the high requirements.Requirements for mass trajectory generation.(3)In order to further improve the estimation accuracy of GLMB filter parameters,this paper abstracts the process of optimizing parameters into the simulation environment of reinforcement learning,and constructs a reinforcement learning algorithm framework.The filter parameters inferred by the LSTM neural network based on the radar point trace data are used as the initial value,and the similarity between the output track of the GLMB filter and the ideal track is used as the reward basis,and the Q-learning and deep Q-based learning are realized respectively.A closed-loop feedback training process for the Value Network(DQN)algorithm.Build an environment through the experimental platform to complete training and evaluation.The experimental results show that both the reinforcement learning framework based on Q-learning and DQN can optimize the GLMB filter parameters.The DQN algorithm has better efficiency and success rate than the Q-learning algorithm,and can meet the needs of high-quality track generation.