THE INSPECTION OF VARIOUS INPUT TRANSFORMATIONS TOWARDS HUMAN DETECTION AND TRACKING

Considering the role of color space in enhancing object detection this paper finds the impact of detection models trained on different color spaces in human tracking system within the Tracking by detection (TBD) framework. A customized dataset over 8k frames including indoor and outdoor human movement videos was developed following the MOT15 structure. YOLO12s was fine-tuned separately on six color spaces RGB, Grayscale, HSV, HSI, YCbCr and YES followed by SORT and DeepSORT for tracking. YOLO12s trained on RGB provides the best MOTA with 31.5% and 50% for SORT and DeepSORT respectively. Competitive results observed from the Grayscale color space.