Research On Video Multi-target Tracking Algorithm Based On Feature Fusion And Data Association

Video multi-target tracking is an active branch in the field of image processing and computer vision,and has a wide range of uses in many fields,such as precision guidance for missiles in the military field;real-time for traffic in the field of intelligent transportation Monitoring and so on.Therefore,multi-target tracking has important practicality and broad prospects for development.However,due to the complexity of the actual application scene and the randomness of the target movement,such as the change of the illumination in the environment,the change of the target size,the similarity between the target color and the background,the moving target is obscured.This paper mainly studies the multi-target tracking from two parts,including multi-feature fusion of video objects and multi-target tracking based on data association.In terms of feature fusion of video objects,firstly,the shortcomings of two feature fusion matching methods are analyzed.On this basis,a new feature fusion method is proposed,which uses three complementary features to perform feature fusion matching.The target is detected by the background difference method,the tracking target is detected,and then the color feature,edge feature and texture feature of the tracking target are extracted.In the case of tracking the change of the illumination of the environment,the feature of the color feature description target is reduced,and the edge feature can not adapt to the target deformation by using the feature of the edge feature to make up the deficiency of the color feature,and the rotation invariance of the texture feature to make up for the lack of edge features,and the establishment of color histogram,edge gradient histogram and texture histogram,the distance between the target histogram and the candidate histogram of each feature is calculated by using the Bhattacharyya distance or the chi-square distance to measure the similarity between the target and the candidate target.And then form a feature fusion matching matrix to improve the matching degreeIn terms of multi-target tracking based on data association,IJPDA(Improve Jointprobabilistic data association)algorithm is proposed,the classical probability data association(PDA)is introduced.The advantage of the algorithm is that the calculation is small and easy to implement.It is robust to single target tracking,but it is prone to misleading for multi-target tracking environment.The JPDA algorithm is suitable for dealing with multi-target tracking,but the number of joint events in the video increases,the number of joint events increases exponentially,that is,the "combinatorial explosion" problem.In this paper,the shortcomings of the JPDA algorithm are improved,and the computational complexity is greatly reduced by screening out the small probability joint events generated by the false alarms.At the same time,the data association can determine the one-to-one correspondence between the measurement and the target,and improve the tracking accuracy...