Detection‐Based Visual Object Tracking Based on Enhanced YOLO‐Lite and LSTM
Aayushi Gautam, Sukhwinder Singh · 2022
Visual object tracking performed using visual trackers is a crucial and standard process inside numerous visual systems. Although it is simple for humans to track objects, visual trackers are still far from their target, requiring them to capture the temporal and spatial relationships between the objects. Many traditional tracking algorithms prove inaccurate, thus demanding a robust and accurate visual tracking approach. We propose a framework that utilizes enhanced YOLO-Lite and LSTM to perform target tracking inside the video to enhance object tracking accuracy. We propose a novel-enhanced YOLO-Lite model, an amalgamation of YOLO-Lite and hybrid spatial pyramid pooling. The module is dedicated to improving object localization by thoroughly utilizing the local and global multiscale feature information inside the video frames. Second, we work with LSTM to obtain the target trajectory for bounding boxes obtained after detection. The entire framework is computationally inexpensive and has the ability to learn historical patterns suitable for efficient and accurate object tracking. The experiments carried out on standard benchmarks VOT-2016, UAV-123, and OTB-2015 validate the efficacy of the proposed framework positively against the futuristic trackers.