Single Object Tracking Based On Deep Learning

Single Object Tracking aims to locate an object given in the first frame in the following part of the video.The single target tracking algorithm has many practical applications in behavior analysis,video surveillance and augmented reality.Although many algorithms have made great progress on this issue in recent years,single object tracking is still a challenging task because of deformation,illumination,rotation,and occlusion.Recently,deep learning methods have greatly improved the algorithm performance in many tasks(such as object classification,object detection,etc.).In this paper,I focus on how to apply the deep learning algorithm to improve the single target tracking algorithm properly.Features in different layers of deep neural networks have different properties: higher layer features contain more semantic information and are more robust to object appearance changes;lower layer features include more spatial information and are more sensitive to the transformation of objects.How to utilize these two features reasonably on the single target tracking task is an interesting problem.Besides,appropriate application of attention mechanism can bring more information in the process of single target tracking,leading to improvement on the performance of the algorithm.Therefore,we propose a multi-level attention mechanism module,which is successfully applied on single object tracking,and has achieved good performance.This multi-level attention mechanism contains a total of four attention models: spatial attention,channel-wise attention,temporal attention,and layer-wise attention.In addition,considering that the existing single object tracking algorithm is in a situation where precision and speed cannot be considered synchronously: regression-based algorithms are mostly fast,but the accuracy is not good enough;algorithms with high precision are mostly based on classification with the very slow speed,which most can only run few frames per second.Therefore,this paper combines the advantages of the two type algorithms and proposes an Actor-Critic single object tracking framework which can balance accuracy and speed.At the same time,the training of the model with reinforcement learning which is famous for its outstanding performance in the problem of sequence decision making is utilized,so that the network can make correct decision output in the single object tracking task.In addition,based on the original Deep Deterministic Policy Gradient,this paper has made some improvements considering the single object tracking task,which makes the network training process converge.The algorithm finally achieves the performance which is better than a lot of non-real-time algorithms,while maintaining speed beyond the real-time requirements(30FPS).