Deep Learning-Based Object Detection for Vehicular Safety: A Comparative Study on COCO, KITTI, and Merged Datasets

Abstract

This study investigates deep learning-based object detection for vehicular safety using the YOLOv8 architecture. The model was tested on three dataset configurations: COCO-only, KITTI-only, and a merged COCO-KITTI dataset. COCO-only training provided broad generalization across many object categories but lacked traffic specialization, while KITTI-only training achieved strong performance in road-centric classes such as cars, trucks, and pedestrians, yet underperformed on general objects. The merged dataset offered a balanced compromise, integrating COCO's diversity with KITTI's domain specificity. Fine-tuning produced limited benefits: on KITTI, it yielded marginal improvements in precision, recall, and mean Average Precision (mAP), while on COCO and the merged dataset the results were neutral or slightly degraded. This suggests fine-tuning is not universally effective and provides value mainly in domain-focused contexts. The model achieved real-time inference at approximately 263 FPS, corresponding to 3.8 milliseconds per image, confirming its potential for embedded and automotive applications such as autonomous driving and advanced driver assistance systems (ADAS).Topic - Artificial Intelligence and its Applications. © 2025 IEEE.