Comparative study of object detection models for automotive in-vehicle environments

Abstract

The rapid progression of autonomous driving and advanced driver-assistance systems has emphasized the need for robust and efficient object detection within vehicle environments. This paper conducts a comparative analysis of state-of-the-art object detection models for in-vehicle environments using public datasets, guided by the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology. The study includes several YOLO models (YOLOv5, YOLOv7, YOLOv8, and YOLOvl0), a transformer-based model (RT-DETR), and a CNN-based model incorporating self-attention mechanisms (YOLO-NAS). The evaluation relies on key metrics such as mean Average Precision (mAP) and inference time to determine each model's effectiveness. The experimental results indicate that while our approach to vehicle interior objects detection for forgotten items shows promise, the overall performance fell short of expectations. The YOLOv7 x variant achieved a recall of 0.632 and an mAP50 of 0.351, offering better precision than YOLOv5 in some cases. However, its 10.3-millisecond GPU inference time underscores the challenge of balancing accuracy and computational efficiency. The findings emphasize the impact of dataset quality and class balance in achieving robust and reliable object detection, highlighting the need for tailored datasets to enhance model reliability in real-world automotive applications. This research contributes to the broader vision of green and smart cities by advancing technologies that improve vehicle safety and support sustainable urban mobility systems.