Deep Reinforcement Learning for Action Based Object Tracking in Video Sequences

Abstract

In this paper, we propose a valuable route for visual object tracker which catches a bounding box to zone of premium physically in the video frames by recognizing the activity got the hang of utilizing the convolution neural systems. The proposed convolution neural network used to control tracking actions is done with various training video sequences and fine-tuned during the actual tracking of the object. Pretrain of the video is done using deep reinforcement learning (RL) along with the supervised learning. Mostly named information from the RL can be utilized for semi supervised learning and assessing through object tracking benchmark dataset, the proposed tracker is confirmed to accomplish a good performance. The proposed method, which operates in real time on without graphics processing unit, outperforms the state of real time trackers with proper accuracy with performance 10%.