Vehicle Distance Estimation Method Based on Monocular Camera

Abstract

Advanced driver assistance system (ADASs) are important on traffic safety. In ADASs, vehicle distance estimation methods can be classified into sensor based, multiple-camera based, and monocular-vision based methods. However, sensor-based methods mainly apply radar information and are sensitive to the interference of buildings. Multiple-camera based methods require more computation loading. Monocular-vision based methods are more practical, however, their performance need to be improved. In this study, we proposed several techniques to improve the accuracy of the monocular-vision based distance estimation. The proposed algorithm is divided into two stages: feature point extraction, and vehicle distance estimation. In feature point extraction, we find the Harris corners and perform road extraction and find the masks by segmenting the road lane and the tire regions according to their colors and relative locations. Then, polygon approximation is applied to get four corners of the lane. After getting critical feature points, we use the geometric relationship between the camera and the tire bottoms to estimate the distance. However, the tilting angle of the camera may highly affect the accuracy of monocular vehicle distance estimation. In practice, the tilting angle is hard to known explicitly. To solve the problem, we adjust the camera angle according to the standard length of road lanes using yellow and blue feature points. Simulations show that the average error of the proposed algorithm is much lower than that of state-of-the-art methods, which indicates the feasibility of the proposed method.