Detection of Lateral Road Obstacles Based on the Haar Cascade Classification Method in Video Surveillance

Abstract

In this paper, we address the problem of improving driving safety on main roads outside built-up areas in sub-Saharan countries, particularly in Senegal, where there are more than 600 deaths per year in traffic accidents. The driver’s field of vision in the rearview mirror, which is limited by the presence of a blind spot, does not always allow him to detect a potential obstacle in time. In this work, we propose a solution for real-time detection of some lateral obstacles on roads outside builtup areas from images acquired by video surveillance. To manage the detection of several obstacles in real time, we used the Haar cascade model to propose a sub-model for each type of obstacle. Thus, these models will be grouped in a single model. To achieve this we first collected data (images). The images collected must be of two types: positive images and negative images. The positive images must contain the obstacles. Negative images are images that do not contain obstacles. The evaluation of our model was highlighted through video image captures made on the road of Kaolack (a region of Senegal) with an accuracy rate of 88% determined from the ROC curve.