Path Aggregation Network with You Only Look Once Version 8 for Real-Time Object Detection in Autonomous Vehicles

Nowadays, the development of autonomous vehicles heavily depends on real-time object detection which allows vehicles to perceive and respond to their surroundings. However, the existing models faced challenges in lighting conditions and real-time scalability due to their high computational complexity which led to result in slower inference times. Hence, this research proposes a Path Aggregation Network with You Only Look Once Version 8 (PANetYOLOv8) for faster inference times and real-time scalability even in varying lighting conditions. Initially, the input data is collected from KITTI and BDD100K datasets which contains high-quality images and realistic driving scenarios by representing environments across different times as well as weather respectively. Then, the input data is preprocessed with normalization and data augmentation to improve image quality. After that, MobileNetv3 is employed to extract the lightweight spatial and semantic features. Finally, the PANet improves feature maps by fusing information across multiple layers in the detection of objects at various scales and YOLOv8 uses these features to quickly predict object classes. From the results, the proposed PANet-YOLOv8 outperformed existing improved YOLOv5 in terms of precision (94.7%) and recall (98.5%) respectively.