Enhancing Object Detection for Autonomous Vehicles Using YOLO-NAS on Bangladeshi Dataset

In autonomous driving, understanding unstructured road scene conditions like in Bangladesh, is crucial as it enables the machine to capture the visual perspective of the environment. Autonomous vehicles depend on a strong object detection model to perceive complex road scenarios and respond to these situations accurately. Recently, Deep Learning (DL) has emerged as a highly promising approach in such object detection tasks. This study explores and evaluates the deep learning-based YOLO-NAS (small, medium, large) algorithm for detecting various objects based on the RSUD20K dataset. The dataset consists of more than 20,000 images, including 130,000 annotations and 13 object categories, and it is tailored to the unorganized and unstructured road conditions of Bangladesh. The algorithm's performance is evaluated by broadly accepted metrics including Precision, Recall, Mean Average Precision (mAP). The findings of this study indicate that YOLO-NAS_M outperforms the other variants with a mean average precision at 50 (mAP@ 50) of 82.6%, closely followed by YOLO-NAS_L at 81.55% and YOLO-NAS_S at 76.4%. Overall, this work demonstrates the capabilities of YOLO-NAS in the context of object detection for autonomous vehicles.