Multi-scale Single Object Tracking Based On Attention Mechanism

Single object tracking is a technology that uses the contextual information in the video to model the appearance and motion information of the target and predicts the target position.It is applied widely in almost all the industries such as video surveillance,visual navigation and military.It usually faces challenges like similar target interference,occlusion,and out of view.With the emergence of large video training datasets,deep learning has provided new opportunities for single object tracking research.It has made notable progress in single object tracking.With the development of deep learning,a large number of single object tracking algorithms have been proposed.The Accurate Tracking by Overlap Maximization(ATOM)algorithm is an excellent single object tracking algorithm.This algorithm is mainly composed of a target classification network for coarse positioning and a target evaluation network for scale estimation.In the process of tracking,the network online learning is used to adapt to target changes.However,in the face of occlusion and similar target interference,the target classification network is easy to classify the target and interference wrongly.In the face of target disappearance,the network online learning stage may lead to model drift.Through the research and analysis of the ATOM algorithm,this thesis proposes a new single object tracking algorithm base on ATOM.The specific improvements are as follows.(1)A multi-scale network for single objcet tracking.Aiming at the difficulty of occlusion and similar target interference,the target classification network in the ATOM has a poor discrimination ability due to the fixed receptive field.A target classification network based on the Inception network is proposed to solve this problem.This network uses asymmetric convolution to increase the multi-scale convolution kernel while reducing the number of parameters as much as possible to ensure that the network can be optimized with a small number of samples during the online learning stage.(2)An online learning algorithm based on the attention mechanism.Aiming at the problem that the online learning algorithm in ATOM cannot adapt to the disappearance of targets,similar target interference and occlusion,this thesis analyzes the changes in the attention map during the tracking.And combines the attention map to design a loss function in the online learning stage,which can enhance the network's ability to discern targets.(3)Sample selection strategy based on attention mechanism.The ATOM algorithm cannot effectively filter the target vanishing samples,the model drift problem may be occured when the target vanishing samples participate in the online learning.A reliable sample selection strategy is proposed to solve this problem.A sample selection strategy is proposed by combining the attention map and the target classification network's score map to select the tracking sample frame.It can retain reliable sample frames and maintain their diversity.This strategy makes the network to quickly learn the characteristics of target appearance and improves the robustness.In order to verify the effectiveness of the method in this thesis,the experimental stage was evaluated on the OTB-50 and OTB-100 test datasets with other classic tracking algorithms.At the same time,this method is compared with other well-performing trackers in the practical application of aerial object tracking to prove its effectiveness.