Streamlined Multi-Object Tracking With Fuzzy Kalman Filtering and Shape-Aware Association Strategies

Abstract

The YOLOv8 family of models excels in object detection and demonstrates competitive performance relative to other methods; however, utilizing these models in resource-limited environments for real-time tracking presents challenges. This study introduces an innovative pedestrian tracking system that integrates a lightweight detector with a two-stage adaptive association mechanism. The YOLOv8 backbone is enhanced using GhostNet modules, resulting in faster inference times and a reduced number of parameters without compromising detection accuracy. Additionally, the ByteTrack association strategy is refined by incorporating shape-aware SIoU (SA-SIoU) for improved spatial alignment and employing a Fuzzy Scaled Adaptive Kalman Filter to better address non-linear motion. Termed FS-ByteTrack, this proposed system has been validated on the MOT17 benchmark, achieving a MOTA of 56.61% and an IDF1 of 65.13%. When compared to the OCSORT baseline, FS-ByteTrack showed a substantial 40.68% increase in processing speed operating at 41.5 frames per second making it a strong candidate for real-time applications requiring accurate tracking in complex scenes.