Research On Object Tracking Algorithms Based On Correlation Filters

Object tracking is an important research direction in computer vision.It is widely used in intelligent surveillance,human-machine interaction,military guidance and other fields.After decades of development,many classical and excellent algorithms have been proposed.However,due to the complexity of environment and the variability of targets,such as illumination variation,deformation,occlusion and so on,object tracking is still a very challenging task.Recently,correlation filters-based tracking algorithms with precision and speed attract extensive attention from researchers.They reduce the time complexity of the algorithms by Fourier transform,and introduce the cyclic shift samples for full training.Thus the algorithms improve the tracking speed and precision effectively.Due to the virtual nature of cyclic shift samples,when the target appears occlusion,motion blur,fast motion and so on,the tracking accuracy is significantly reduced.In addition,a fixed update rate and a linear update mode would lead to erroneous updates or make algorithms unable to adapt to the variance of target appearance in time.Aiming at the shortcomings of correlation filters-based tracking algorithms,this article proposes improvements and design experiments to verify them.The main work of this article is summarized as follows.(1)A tracking algorithm based kernel correlation filters and sparse prototypes is proposed.In the detection stage,reasonable candidate samples are selected,real response values are calculated,and sparse prototypes are combined to locate targets,so as to reduce the negative effects of cyclic shift samples.To adapt to the change of the target appearance and avoid the pollution of the model,the update mode is selected according to the noise detection results and the adaptive update rate is defined according to the tracking confidence.Experiments show that the algorithm can effectively deal with the problems of occlusion,fast motion and motion blur,and the tracking performance is improved obviously.(2)A correlation filters tracking algorithm with multi-strategy combination is proposed.Based on the correlation filters with multiple features,it uses a variety of features to describe the target to enhance the tracking stability in complex environments.Then it judges the tracking status by combining tracking confidence with target similarity.The tracker and target sample library will update when the tracking status is normal.The algorithm will expand the detection scope and use a multi-layer filtering mechanism to relocate the target when the tracking state is abnormal.Experiments show that the algorithm is robust to many challenging factors in tracking process,and it can recover the target in time when the target is lost.(3)The above two algorithms are improved by scale filters.By adapting to the changes of the target scale,more effective target information can be obtained and the interference of the surrounding background information can be reduced.Thus the update accuracy of the algorithm can be guaranteed,and the tracking performance of the algorithm can be further improved.Experiments show that the improved algorithms can adapt to the scale change of the target effectively,and achieve better tracking performance on the multi-scale data than the original algorithms.