Pedestrian tracking based on deep learning

• 主要作者: Xia, Yuqi
• 其他作者: Yap Kim Hui
• 格式: Thesis-Master by Coursework
• 語言: English
• 出版: Nanyang Technological University 2024
• 主題: Computer and Information Science; Deep learning; Multiple object tracking
• 在線閱讀: https://hdl.handle.net/10356/176521
• 機構: Nanyang Technological University

The methodology of pedestrian tracking extensively incorporates techniques derived from the field of computer vision. This domain is primarily concerned with the processing and interpretation of visual data obtained from still images or video sequences. Tracking pedestrians simultaneously requires advanced algorithms capable of detecting, identifying, and continuously monitoring these subjects through successive video frames. The original FairMOT algorithm, conceived as an innovative model, is designed for precise multi-object tracking. It operates using a dual-branch structure: one branch predicts pixel-wise objectness scores, while the other captures features for re-identification. The primary goal of this model is to balance the tasks effectively, aiming for superior tracking and detection accuracy. Despite its strengths, the algorithm encounters challenges with inaccurate target detection, often exacerbated by frequent target occlusions. Our research is dedicated to optimizing the computational efficiency of the tracking algorithm and significantly enhancing its tracking accuracy. This dissertation presents two enhancements to the FairMOT framework, improving pedestrian tracking accuracy. It explores integrating the Cross-Stage-Partial (CSP) network structure into CenterNet’s backbone network and introduces a threshold-based method for more effective Intersection over Union (IoU) matching detection. These enhancements result in a 0.2% increase in detection accuracy over the standard FairMOT algorithm, significant improvements in HOTA (2.60%), MOTA (0.56%), and IDF1 (3.7%) scores, and a 0.64% accuracy gain over DeepSORT and Track-RCNN algorithms.