Weak Target Detection And Tracking Based On Particle Filter

Target detection and tracking has been widely used in the military,electric power system fault detection and our daily life.Interframe Difference algorithm,Background Difference algorithm,Optical Flow method,Mean Shift method and Kalman Filter etc.are the common methods of target detection and tracking.However,when the weak target is under complex background,these algorithms are limited.The background of the weak target is always dynamic and the noise of the background is strong,so the Background Difference algorithm and the Interframe Difference algorithm are not appropriate.The Optical Flow method has high sensitivity of noise,occlusion,light and transparency,it is difficult to calculate the optical flow field.Mean Shift method is not able to track low SNR target,if the target is under occlusion,the tracking will be not correct.Kalman Filter(KF)method can be used in the Linear and Gaussian system.However,the actual backgrounds always have obvious characteristic of non-Linear and non-Gauss.Particle Filter(PF)is a sequential Monte Carlo method which is based on a recursive Bayesian posterior probability theory.Posterior probability distribution of the system state can be replaced by a series of random particles,so the target under non-Linear and non-Gaussian circumstance can be tracked.However,in the process of traditional PF algorithm,new observation information is not used in the importance distribution function,which leads to inaccuracy.The Unscented Particle Filter(UPF)algorithm is combined of Unscented Kalman Filter(UKF)algorithm and traditional PF algorithm,the new observed value of the state is used in the importance sampling.However,the ultraviolet corona video is under low Signal Noise Ratio(SNR),the gray information is the only feature.Moreover,the position of the target can not be observed in a single frame.Therefore,the tradition UPF cannot be satisfied to the background of this paper.Considering the weights of the particles are not accurate enough because of the lacking information.The UPF algorithm is improved as follows:extracts the candidate target regions(the target and the noises)of the binary image,limits velocity of the target,then calculates the weights of the particles by comparing the gray histograms.In addition,if the target is under occlusion,this paper put forward a solution.The research results show that the improved algorithm is better than the other two algorithms.Last,this papar analyzes the shortcomings of the improved UPF algorithm,makes a summary and prospects of the future work.