Depth-aware object detection and region filtering for autonomous vehicles: a monocular camera-based novel integration of MiDaS and YOLO for complex road scenarios with irregular traffic

Abstract

In this era of technological advancement, autonomous vehicles strive towards revolutionizing the transportation system. Object detection and distance measurement are critical for autonomous driving, particularly in dynamic and unpredictable environments. This paper presents a new approach for improving object detection and depth estimation in autonomous vehicles, with a focus on the complex road conditions of Bangladesh. Using MiDaS for depth estimation and YOLOv8 for object detection, we introduced the Depth Aware ROI Filter (DARF) to refine the region of interest, enhancing detection accuracy while reducing processing time by more than 85%. By focusing solely on depth-relevant areas, our method optimizes object detection, combining depth information and object classification essential for autonomous navigation. Tested on a diverse dataset of Bangladeshi road scenarios, including various vehicles and weather conditions, our approach ensures real-world applicability. Additionally, we explored several regression models to calculate the relative depth-to-real distance scale factor, further improving depth accuracy. This work demonstrates the potential of our system to enhance real-time decision-making in autonomous vehicles, paving the way for advancements in object tracking and system optimization in complex road environments.