Development of Naïve Algorithm to Identify Objects on Road and Measuring Distance using RANSAC Algorithm

Abstract

Depth or distance estimation approaches that rely on deep learning need a substantial volume of data, and maintaining domain invariance is a hurdle. Hence, this study presents a streamlined and efficient method called the single view geometric approach. This technique utilizes the geometric characteristics of the road lane markers to accurately estimate distances. It seamlessly interacts with the lane and vehicle recognition components of an already established Advanced Driver Assistance System (ADAS). Our technique incorporates innovation in two aspects: firstly, it utilizes cross-ratios of lane borders to estimate the horizon. (2) It calculates an Inverse Perspective Mapping (IPM) and camera elevation based on a given lane width and the identified horizon. The distances of the cars on the road are determined by projecting the image point back onto a ray that intersects the rebuilt road plane. During the assessment process, we used data as the reference standard and assessed the effectiveness of our system in comparison to image dataser and the most advanced deep learning based monocular depth prediction methods. The results from the evaluation on three publicly available datasets shown that the suggested approach consistently maintains a Root Mean Square Error ranging from 6.10 to 7.31. It demonstrates superior performance compared to other algorithms on two of the datasets, but lags behind one deep learning approach on KITTI.