Pedestrian detection and hybrid tracking framework for occluded scenarios

Pedestrian tracking is an essential aspect of Intelligent Transportation Systems (ITS). The current tracking system faces challenges in a complex traffic environment that entails many pedestrians, occlusions, irregular motion, and unstructured roads, making it difficult to detect and track pedestrians. Therefore, to overcome these challenges, this study proposes a tracking-by-detection approach that integrates detection and tracking: YOLOv11n for detection and a hybrid tracker that combines BoT-SORT and DeepSORT. The hybrid tracker is based on motion prediction and re-identification, ensuring continuous identity tracking of pedestrians even in complex traffic environments. The experiment evaluated the performance of different YOLO models and conventional models, with YOLOv11n detection showing a strong compromise between precision and speed, achieving 86.0% mAP@50 and 55 FPS. When used for tracking, the hybrid tracker achieves MOTA 76.8% and IDF1 75.2%. In addition, the hybrid tracker reduces identity switches to around 180 and recovers around 4200 bounding boxes that are occluded in the IDD Pedestrian dataset. The model has achieved a good compromise when it comes to the trajectory estimation, with an average shift of 1.42 pixels. This result indicates that the proposed system exhibits better tracking performance in congested and unstructured traffic settings.