Object Detection In Autonomous Vehicles

Authors: Rohan Chaudhary, Mayank Chauhan, Dr. Partap Singh

Abstract: In the context of autonomous vehicle (AV) perception, this study examines the implementation and performance assessment of the You Only Look Once version 5 (YOLOv5) model for object detection (OD). Safe and reliable navigation requires high-throughput algorithms, which require accurate and real-time detection. Roboflow's carefully selected dataset of approximately 15,000 photos was used in the study, with a focus on specific classes like cars, trucks, pedestrians, traffic lights, and bikes. PyTorch was used to train the YOLOv5 architecture for 100 epochs to enhance generalization across different lighting and angle conditions, and robust data augmentation techniques like brightness adjustment and scaling were included. The model's efficiency is attributed to its single-stage detection approach, which places emphasis on the necessary inference speed for real-time decision-making in autonomous systems. The trained model achieved a mean average precision (textmAP) of around 85%, and both precision and recall metrics exceeded 80% in terms of numbers. Critical analysis revealed that detection accuracy was significantly higher for large, frequently occurring classes (cars) than for smaller, safety-critical classes (pedestrians and traffic lights), indicating a performance gap. Further architectural improvements are required due to the structural limitation caused by class reduction and scale variations. The paper concludes with a roadmap for improved robustness, which includes migrating to sophisticated architectures like YOLOv8 or Detection Transformers (DETR), to achieve near-perfect recall rates for high-consistency objects.

DOI: http://doi.org/10.5281/zenodo.21418279