Research Of Adaptive Object Tracking Based On Fusion Of Multilayer Convolutional Features

Visual object tracking is one of the research hotspots in the field of computer vision.Although many theories and algorithms have been proposed for object tracking,it is still a challenging problem to effectively improve the robustness of the object tracking algorithm because of the complex and diverse application scenarios.Discriminative scale space tracking(DSST)algorithm has been widely applied as a mainstream tracking algorithm.However,because the artificial features used by DSST cannot capture the semantic information of the target,the ability to describe the target is difficult to meet the changes in the complex environment during the tracking process,so the problem of object tracking failure occurs easily in complex scenarios.This thesis proposes to use the convolutional neural network's multilayer convolutional features to improve the robustness of DSST object tracking algorithm.The main works are summarized as follows:1.The basic theory of convolutional neural network and feature extraction based on model VGG-19 are studied.The hierarchical features extracted based on model VGG-19 are analyzed in depth.In the network,the features of each layer reflect different information components.The low layer feature maps with high resolution contains more detailed information,while the high layer feature maps have low resolution but can extract more semantic information.Through the experimental analysis,the optimal convolution layer feature response maps are selected for fusion in order to obtain more effective feature expression capabilities2.An adaptive correlation filter tracking algorithm based on multilayer convolutional features fusion is proposed.By combining the scale space correlation filtering algorithm with the multilayer convolutional features,the target position is estimated based on the multilayer convolutional features,the target scale is estimated based on the HOG features.A model adaptive updating mechanism is introduced to evaluate the tracking confidence by calculating the peak sidelobe ratio(PSR),and determine whether the location filter model is updated,then achieve the adaptive tracking of the object.The experimental results indicate that the proposed algorithm has better adaptability to the tracking tasks in complex scenes,and shows more stable and accurate tracking performance compared with the other latest mainstream tracking algorithms.