Research On Box-particle Filter Based Multi-target Visual Tracking

Visual tracking can be described as a problem of estimating state parameters of targets of interest like position,velocity,size etc in different frames of a video.Only achieve an accurate track of the target can continue to study more complex visual tasks.As a frontier research topic in the field of computer vision,visual tracking has huge development space in intelligent monitoring,human-computer interaction etc.In a real tracking scenario,the number of targets is time-varying,the state of the target is also uncertain,the corresponding multi-target tracking technology for solving these problems has been developed.In recent years,the multi-target tracking technology based on RFS theory which avoids the complex data association,is suitable for dealing with multi-target tracking problems,and has high computation efficiency,suitable for engineering.This thesis mainly improves multi-target tracking algorithms based on RFS in visual tracking,including improving the tracking efficiency,appending trajectory recognition and building reliable likelihood model.Firstly,the basic theory of box particles is introduced.Box particle filter is a generalized particle filter which combined interval analysis with Monte Carlo algorithm to achieve a method of dealing with non-precision measurements.Compared with the traditional particle filter,it ensures the tracking accuracy with a small number of particle and computation costs.Thus the operating efficiency is improved to satisfy the requirement of real-time video tracking system.Secondly,multi-target visual tracking method based on box particle filtering and RFS is studied deeply.A multi-target tracking method of box particle PHD with trajectory recognition is proposed.It is updated by the position provided by the target detection algorithm.In order to achieve good tracking effect,a fast target detection algorithm is designed to produce a more accurate position measurement.It achieves an adaptive threshold setting in background difference based on Gaussian initialization strategy,and improves the adaptability of the algorithm to the environment.The algorithm of trajectory recognition is designed to distinguish the target track by using the characteristics of the target,and further eliminate the clutter in the target state sets which ensure the tracking accuracy.Finally,to overcome the shortcomings of lower degree of descriptiveness only using single position feature,an improved likelihood model box particle PHD visual tracking algorithm is proposed.Using the position and SIFT feature describing the target,this algorithm compensates the insufficiency of the traditional box particle PHD filter only using the movement characteristics and ignoring the apparent characteristics of the target.Experiments show that the proposed algorithm can achieve good performance in multi-target visual tracking in complex situations such as target merge,split,occlusion etc.