DeepSCT: Deep Learning Based Self Correcting Object Tracking Mechanism
Khush Agrawal, Rohit Lal, Himanshu Patil, Surender Kannaiyan, Deep Gupta · 2021
This paper presents a novel mechanism, DeepSCT, to handle the long-term object tracking problem in Computer Vision. The paper builds around the premise that the classical tracking algorithms can handle short-term tracking problems efficiently; however, they fail in the case of long-term tracking due to several environmental disturbances like occlusion and out-of-frame going targets. The relatively newer Deep Learning based trackers have higher efficacy but suffer from working in real-time on low-end hardware. We try to fuse the two methods in a unique way such that the resulting algorithm has higher efficiency and accuracy simultaneously. We present a modular mechanism, which can accommodate improvements in its sub-blocks. The algorithm was tested on the VisDrone-SOT2019 dataset for a person tracking task. We quantitatively and qualitatively show that DeepSCT significantly improved classical algorithms' performance in short-term and long-term tracking problems.