Multi-target Tracking Based On Data Association And Trajectory Evaluation

In the field of image processing and pattern recognition,Multi-target tracking technology due to its wide application prospects,largely social demand,it is becoming more and more hot topic by researchers.This paper mainly research on two major aspects of target detection and multi-target association.A robust tracking solution is proposed to solve the difficulty of illumination variation? mutual occlusion and label switch in the process of object tracking.A target detection method based on symbiotic features and gradient self-similar features are firstly introduced.Firstly,the HOG feature and LBP feature of each frame are extracted,the pairwise gradient self-similarity GSS feature between the local gradient blocks are calculated according to the HOG features,and the Co LBP features are obtained according to LBP features;Secondly,Using the FGM method to remove information independent GSS features,and the discriminate DGSS features are further obtained;Finally,the two-level cascade classifier is used to evaluate the performance of pedestrian detection.The first classifier uses the linear SVM classifier based on HOG-CoLBP feature training to remove most of the easily easy negative samples;for the second classifier,taking into account the HOG feature is the base of generating GSS characteristics,so we reuse the calculated HOG features to generate the corresponding GSS features,then the Real-AdaBoost classifier based on HOG-GSS feature training is used to detect the targets in the candidate image area,and complete pedestrian detection is realized.it is demonstrate from the experimental results that proposed method can extract the foreground target in the case of illumination and dynamic disturbance.The paper focuses on a multi-target tracking algorithm based on data association and trajectory evaluation.First,the data association and trajectory evaluation are integrated into the same conditional random field model(Conditional Random Field,CRF),which is transformed to obtain the minimum energy problem.Secondly,if two of the same tracking label appear in a physical space from time to time,the symbiotic label cost is used to constrain the correlation;in addition,the pairwise energy term between the different observation targets from the space is introduced to prevent the occurrence of false data label.Finally,in the process of energy optimization,this paper uses the improved ?-expansion algorithm and the gradient descent method to solve the minimum energy in non-convex and non-sub module functions.The experimental results of PETS2009/2010 benchmark and TUD-Stadtmitte videos sequence database show that the proposed algorithm is superior to the state of art of multi-target tracking technology.