Performance Evaluation of YOLO Models in Varying Conditions: A Study on Object Detection and Tracking

Object detection and tracking are essential components in autonomous systems, enabling accurate environmental perception and real-time decision-making. Despite the widespread use of advanced frameworks such as YOLO, RCNN, SSD, FPN, and RetinaNet in applications like autonomous driving and surveillance, these systems often experience degraded performance in images with poor lighting conditions caused by adverse weather and other external factors. Recent developments have introduced various deep-learning and image-processing techniques to address these challenges. This paper provides an extensive evaluation of three state-of-the-art object detection models-YOLOv5, YOLOv8, and YOLOv9 under diverse experimental conditions, including varying confidence thresholds (50%, 60%, 70%, and 80%), different object poses, lighting conditions (day and night), and camera configurations (static and dynamic). The evaluation is conducted using key performance metrics such as mean Average Precision (µAP), Intersection over Union (IoU), Multiple Object Tracking Accuracy (MOTA), Multiple Object Tracking Precision (MOTP), Identity F1 score (IDF1), precision, and recall. The study offers a comprehensive analysis of the models' performance and suitability for different real-world applications, delivering practical insights for model selection and deployment. Additionally, it highlights potential areas for future research aimed at enhancing model robustness and effectiveness in challenging environments.