Research On Probability Hypothesis Density Filter Multi-target Tracking Algorithm Based On Variational Bayesian Inference

Since the number of tracking objects in the multi-target tracking system is no longer single at the same time,it is fundamentally different from the single-target tracking,and it is more in line with practical engineering applications,such as civilian monitoring or enemy reconnaissance in the military field.Nowadays,facing the increasingly complex and changeable tracking background,especially the tracking targets may appear derivation,rebirth and disappearance at any time.It is also accompanied by the constraints of the sensor’s own conditions make the target missed detection and interference caused by clutter in the environment.These unfavorable factors will limit the estimation accuracy of the filter output target state,so solving these problems has always been a hot issue for scholars to study the direction of multi-target tracking.In the multi-target tracking algorithm based on random finite set,the multi-target state values and all measurement values are respectively constructed into sets,which participate in the prediction and update of target state in the form of random sets,so as to avoid the complex association resulting in too much calculation and affecting the real-time performance of the system.Based on the random finite set theory,the probability hypothesis density filter proposed to approximately obtain the first-order moment of the posterior probability of the target,so that it can acquire an approximate solution in the application of multi-target system.The filter is usually based on the assumption that noise follows Gaussian distribution,but this condition is often not consistent with real tracking scenarios.Since the Gaussian property of noise will be destroyed when the sensor is subjected to electromagnetic interference or the model cannot be tracked due to high maneuvering of the target.Therefore,the improved method proposed in this paper is mainly to solve the problem that the tracking accuracy of the filter is reduced due to such phenomena.Considering that the process noise and measurement noise are unknown and time-varying,which will greatly degrade the tracking performance of the filter,the strong tracking PHD filter based on variational Bayesian is proposed to reduce the interference of this problem to the filter.Firstly,the variational Bayesian inference method is mainly used to model the inaccurate measurement noise as the inverse wishart distribution to approximate the real distribution.Secondly,the strong tracking principle is introduced,and the fading factor contained in the principle is used to correct the state error covariance generated by the inaccurate process noise covariance in real time.Finally,the measurement noise is extended to the augmented state of the single target to form the joint distribution of Gaussian Inverse Wishart(GIW),which constitutes the GIW implementation of the probability hypothesis density filter.The adjusted parameters are involved in the variational Bayesian fixed-point iterative cycle,which mainly outputs the target estimated state value and the corresponding error covariance.This paper also designs two experimental scenarios,which are the different values of the key parameters and the increase of the number of targets in the time stage,so as to evaluate the robustness of the filter.The simulation results show that the proposed algorithm has higher tracking accuracy than the comparison algorithm under the same experimental conditions.For the tracking environment,the process noise and measurement noise have abnormal values due to the sensor’s own reasons or the dynamic model deviation,so the filter has biased estimation of the state of multiple targets.Therefore,this paper designs a gaussian-student’s t mixture distribution PHD robust filtering.Both process noise and measurement noise contain outliers,which usually lead to heavy-tailed noise,and the noise is also accompanied by non-stationary for random outliers.Therefore,the proposed algorithm first constructs the noise through Gaussian Student’s t Mixture Distribution.Then the measurement noise covariance and prediction state error covariance in the mixed distribution component are iteratively deduced in the form of variational Bayesian approximation.Finally,the adjusted parameters are used to update the target estimation value,state error covariance and target weight value until the set accuracy standard is reached.In order to evaluate the tracking accuracy of the proposed filter when abnormal values appear in the noise,two cases are designed.The outliers only appear in the measurement noise,and they appear in the measurement noise and exist in the process noise.The different probabilities of abnormal values are set for each scene.Combined with the experimental results,the tracking error of the proposed filter under the same conditions in these two cases is more advantageous than that of the comparison algorithms.