Car detection based on road direction on traffic surveillance image

Abstract

One of the challenges in car detection is to be able to detect car in any viewpoint from traffic surveillance camera. The size, shape, and appearance of car are different if they are viewed from various viewpoint of traffic surveillance cameras. Car also has the most variety of models compared to the other vehicle. However, car poses usually follow road direction. Therefore, this research proposes a method to detect car based on road direction. The method utilizes 3D car models to generate car poses, groups them into four pairs of viewpoint orientation: 1) front / back view, 2) top left / bottom right view, 3) top right / bottom left view, and 4) left side / right side view, then builds car detectors corresponding to each orientation. On traffic surveillance image, road area is extracted to localize the detection area and road direction is estimated to determine the car detector that will be used by Linear-Support Vector Machine (Linear-SVM). Finally, SVM classifies the features extracted by Histogram of Oriented Gradients (HOG) to detect cars. The test result on various viewpoints of traffic surveillance image gives 0.9098 of Balance Accuracy (BAC).