GS-YoloNet a lightweight network for detection, tracking, and distance estimation on highways

The perception component is a critical element within the autonomous driving system, a complex system that requires careful consideration. Current research in perception for autonomous driving predominantly focuses on identifying vehicles, lanes, and traffic signs, while overlooking other potential factors contributing to traffic accidents. Notably, many highway accidents are caused by the presence of wild or wandering animals. To address this gap in knowledge, our study involved the creation of a dataset comprising 1050 images of large animals that could potentially be encountered on highways. Additionally, we proposed an enhanced Yolo model by modifying its architecture, specifically by replacing the C3 module with C3Ghost. This modification resulted in a reduction of parameters to less than 3.7 million, representing only 52.7% of Yolov5s, while achieving an average accuracy exceeding 95% for each animal type (mAP% 0.5). Furthermore, our GhostSort-YoloNet (GS-YoloNet) integrates the Deep Sort algorithm to enable real-time tracking and speed assessment of multiple targets, demonstrating significant practical utility.