Enhance multi-object tracking with learnable re-identification

• Main Author: Hu, Zihao
• Other Authors: Lin Zhiping
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Computer vision; Multi-object tracking
• Online Access: https://hdl.handle.net/10356/176709
• Institution: Nanyang Technological University

Multi-Object Tracking (MOT) has a broad range of applications in various domains, including video surveillance, autonomous driving, and healthcare monitoring. Despite advancements in MOT algorithms, challenges persist, especially in handling identity switches caused by occlusions and other factors. Re-identification (re-ID) techniques have been widely adopted to associate tracks with detections. Nevertheless, re-ID often suffer from the lack of real-time adaptability. To address these limitations, this project introduces a novel approach, learnable Re-Identification, aimed at enhancing MOT performance. An exhaustive review of existing literature and methodologies was conducted to compare relevant MOT algorithms to identify a suitable model as the baseline model of the project. To be more specific, FairMOT is chosen as the baseline model based on its Multi-Object Tracking Accuracy of 69.8% and Identification F1 score of 69.9% on the test set of MOT17 with MOT17 train set as the training data. Some popular datasets widely used in the task of MOT were scrutinized as well. The need for an adaptive re-ID solution capable of extracting task-specific features in complex MOT scenarios was elaborated and the project developed a learnable module to mitigate the issues of adaptability of existing MOT algorithms, enabling real-time adaptation to evolving tracking challenges. The project concludes that the proposed learnable re-ID network with optimized hyperparameters can improve the performance of baseline FairMOT on several video sequences of MOT17 dataset despite some limitations. The idea of learnable re-ID is noteworthy and deserves to be further studied to form the new paradigm of Multi-Object Tracking.