This project is a study on Multi-Object Tracking (MOT) based on TransTrack. It examines the robustness of TransTrack under various adverse conditions such as poor lighting, motion blur, and snow.
- Why TransTrack?: Investigates the advantages of using a transformer-based architecture for MOT.
- Performance Evaluation: Evaluates performance in challenging environments and highlights the strengths and limitations of TransTrack.
- Dataset: We augmented the MOT17 dataset and used transfer learning to improve tracking accuracy.
- Results:
- Strengths: Effective tracking under diverse real-world conditions.
- Limitations: Challenges under adverse conditions.
These insights aim to provide directions for future improvements in multi-object tracking systems.
The architecture of TransTrack uses a transformer-based model with two decoders: one for detecting new objects and another for tracking previously detected ones. This helps in enhancing tracking accuracy across frames, especially under occlusion and changing scenarios.
The project also focused on assessing the performance of TransTrack in environments with:
- Lighting Effects
- Snow Effects
- Motion Blur
Performance metrics such as MOTA and IDP recall were analyzed before and after applying transfer learning techniques.
The entire development pipeline used in the project is represented below:
The metrics to evaluate performance include:
- IDP Recall
- MOTA (Multi-Object Tracking Accuracy)
- IDs Tracking
After applying transfer learning, the model's performance was compared against its initial results to illustrate the impact.
- G. Ciaparrone, F. Luque Sánchez, S. Tabik, L. Troiano, R. Tagliaferri, and F. Herrera, “Deep learning in video multi-object tracking: A survey,” Neurocomputing, vol. 381, pp. 61-88, 2020.
- P. Sun, J. Cao, Y. Jiang, R. Zhang, E. Xie, Z. Yuan, C. Wang, and P. Luo, “TransTrack: Multiple-Object Tracking with Transformer,” arXiv preprint arXiv:2012.15460, 2020.