A Hybrid System for Object Detection and Distance Estimate Using a Monocular Camera Designed for Autonomous Vehicles

Abstract

This paper proposes a sensing system for an autonomous vehicle for detecting surrounding objects and for identifying their type. This system uses depth maps estimated from RGB images captured by a monocular camera to calculate their distance from the vehicle to minimize cost. The system uses a pre-trained deep learning model called YOLO v8, which has been fine-tuned on the Kitti dataset to detect and recognize objects obstructing the vehicle’s path. The Depth maps for images captured from the monocular camera can be estimated using a deep-learning model called AdaBins. Finally, the system combines the resulting information from both previous models to estimate the distances of objects in real-time. The proposed system achieves a mean Average Precision (mAP) of 74%, an average object detection time of 8.5 milliseconds whaen tested on a video clip with a frame rate of 30 fps, and a relative error in estimating the object’s distance ranging between 4% and 20% depending on its position within the camera’s field of view.