A Modified Lightweight DeepSORT Variant for Vehicle Tracking

Object tracking plays a pivotal role in Intelligent Transportation Systems (ITS), enabling applications such as traffic monitoring, congestion management, and enhancing road safety in urban environments. However, existing object tracking algorithms like DeepSORT are computationally intensive, which hinders their deployment on resource-constrained edge devices essential for distributed ITS solutions. Urban mobility challenges necessitate efficient and accurate vehicle tracking to ensure smooth traffic flow and reduce accidents. In this paper, we present a modified lightweight variant of the DeepSORT algorithm tailored for vehicle tracking in traffic surveillance systems. By leveraging multi-dimensional features extracted directly from YOLOv5 detections, our approach eliminates the need for an additional convolutional neural network (CNN) descriptor and reduces computational overhead. Experiments on real-world traffic surveillance data demonstrate that our method reduces tracking time to 25.29% of that required by DeepSORT, with only a minimal increase over the simpler SORT algorithm. Additionally, it maintains low error rates between 0.43% and 1.69% in challenging urban scenarios. Our lightweight solution facilitates efficient and accurate vehicle tracking on edge devices, contributing to more effective ITS deployments and improved road safety.

@article{El-alami2024,
  title     = {A Modified Lightweight DeepSORT Variant for Vehicle Tracking},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {10},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Ayoub El-alami and Younes Nadir and Khalifa Mansouri},
  doi       = {10.14569/IJACSA.2024.0151067},
  url       = {https://doi.org/10.14569/IJACSA.2024.0151067}
}