Obstacle Detection Based On Monocular Vision Research

In the last few years, vision-based driving assistance systems and automation technologies in Intelligent Transportation Systems (ITS) have been developed rapidly for realizing safe driving to prevent car accidents. In these systems, the ability to detect obstacles and warn the drivers of the presence of obstacles to help them take action in advance is essential. The objective of this thesis is to propose an obstacle detection system. This system processes video data captured by a single monocular camera mounted on the vehicle and outputs the shapes of obstacles.In the first place, Gaussian mixture model (GMM) is a technique for reconstructing the background and is effective for extracting moving objects when camera is static. But according to the characteristics of a road, we can also employ GMM to estimate the background image in the case of a moving camera. Once the background is obtained, all objects (either static or moving objects) on the road can be extracted by comparing the input image with the background image. In the second place, we use two consecutive image frames, and warps the first image according to the geometrical relationship between these two images. The road region is then extracted by comparing the warped image with the second image. In the third place, non-road region is divided into obstacle region and noise region. At last, using these three kinds of regions (road region, obstacle region and noise region), we can delete all things which are not obstacles in the foreground images.This system is examined using experimental videos captured both in artificial scenes and in true scenes. In the performed experiments, it is shown that the proposed system detects both static and moving obstacles such as pedestrians and boxes on a road successfully.