Object Tracking Based On Genetic Algorithm And State Transition

Computer vision(CV)uses cameras and computers to replace human eyes and human vision system to achieve automatic collection,analysis,processing and response to the external world’s visual information.CV can make the computer understand the real world,and it has become one of the most active subjects in the field of artificial intelligence.Although object tracking technology has been developed for many years and has made great progress and application,but because of the complexity of the real environment and the changeable tracking target,it is still a great challenge to develop a generalization tracker with high accuracy,robustness and real-time performance.In order to cope with these challenges,the main innovation of this article is as follows:(1)Aiming at the challenge of single object tracking,this paper proposes visual tracking with genetic algorithm augmented logistic regression.It has been innovatively improved from four aspects:motion model,appearance model,mechanism of template update,and selection mechanism of positive and negative sample and so on.First,in order to solve the problem of candidate area selection,a novel intelligent motion model is proposed in this paper,it not only uses particle filter algorithm to do local search,but also creatively uses genetic algorithm to do global search,which effectively solves the problem of fast motion of target;Then,a novel model updating mechanism is proposed to deal with the effect of the target deformation in real time;Then,a robust positive and negative sample selection mechanism is proposed to effectively solve the problem of tracking drift;Finally,it is effective to combine the FHOG features with the Lab features as the appearance of the target,and further enhance the robustness of the tracker.Experiments and analysis of tracker components on OTB demonstrate the effectiveness of the algorithm innovation,and summarize the important factors that affect the overall performance of tracking algorithm.The overall performance comparison between the 9 trackers in the CVPR2013 benchmark[15]and 7 recent performance trackers is made,which proves that the tracker can generate the best tracking performance.(2)Aiming at the challenge of multi target tracking,a multi-target tracking algorithm based on state transformation is improved.First of all,a multi target tracking framework[35]based on state transition is proposed in this paper.The framework has good scalability and generalization ability,and other efficient detection algorithms and tracking algorithms are easy to transplant;Then,this paper transforms the complex multi target tracking problem into a decision conversion problem and a corresponding data association problem between five simple tracking states;Moreover,the generalization performance and data association ability of the decision function are further enhanced by combining the advantages of two methods of offline learning and online learning.Finally,two sets of experiments are done in the multi target tracking benchmark database[70]to verify the performance of the tracker.Meanwhile,9 tracking devices with excellent performance in multiple tracking domains are compared with experiments,which proves the robustness of the tracker.