Abstract

The rapid growth in autonomous vehicle technology has demanded accurate and efficient object detection systems. The traditional deep learning models have delivered strong detection accuracy; however, they often required heavy computational resources, which made real- time deployment difficult in embedded automotive platforms. The gap between speed and accuracy has created a challenge, especially in dynamic driving environments where the detection delay may risk safety. This study investigated a lightweight real-time detection framework based on improved YOLO variants optimized for low-power environments. The method used knowledge distillation, structure pruning, and feature compression to reduce redundant layers that which do not contribute to final prediction accuracy. A quantization- aware training approach was integrated to enhance efficiency on embedded hardware. Transfer learning was adopted using a pre- trained YOLOv5s backbone, followed by fine-tuning using an annotated autonomous driving dataset. The experimental results indicate that the proposed lightweight model achieves faster inference with higher accuracy. The optimized network processes live video frames at 47 FPS and maintains a mean average precision of 95 percent. The model records a precision of 96 percent and a recall of 94 percent, which surpasses Faster R-CNN, SSD, and Tiny-YOLO baselines. The inference time reduces to 19 ms on embedded hardware, which confirms suitability for real-time autonomous driving perception.

Authors

Belwin J. Brearley¹, K. Regin Bose², N. Kanagavalli³
B.S. Abdur Rahman Crescent Institute of Science and Technology, India¹, Rajalakshmi Institute of Technology, India^2,3

Keywords

YOLO, Autonomous Vehicles, Object Detection, Lightweight Model, Real-Time Detection