Research Of Vehicle's Drivable Road Region Detection Based On Monocular Vision

In recent years, with the general popularity of the motor vehicle and increase of the private cars, traffic safety has become the focus of global world. Because of the complexity and diversity of information on the road conditions, the driver's attention can not keep focused, resulting in missing data of traffic and causing serious traffic accidents. Therefore, the road region detection technology, as one of the core technology of intelligent vehicles has been more popular. As an important part of the driver assistance systems and unmanned systems, the research of road region detection technology can not only help the driver access to relevant information in the scene area of the road but also be effectively guaranteed for other driver assistance technologies (such as lane detection, collision warning of obstacles etc.). So road region detection technology is very necessary. In this paper, combining with the theory of applied mathematics, we set up mathematical models and propose a kind of road region detection algorithm based on monocular vision.First of all, based on the assumption that the road region is flat, in this paper we propose principle of homography, and we detect road region by the principle that the road region is overlapped for the image transferred by homography matrix of the previous frame and after. Firstly, to obtain homography matrix, we need feature points matched on the road region, so we propose a pre-identification algorithm based on SVM (support vector machine) with LBP and HOG features and extract features of image block divided to do road identification. Secondly, we use SURF and Flann algorithm to select and match feature points of the road, thus we obtain the homography matrix according to the matching points, and then we get the final homography matrix after optimization by RANSAC. Finally, according to the obtained homography matrix we make transformation to the image and set the size of image's window block and the threshold for road region detection by single frame (i.e. frame-frame).Then, do multi-frame information fusion by combining the current frame and the historical frame information, we proposes a probability updating model for the fusion of historical frame's detection information, thus we complete the multi-frame fusion detection of road region in the image.The last, by the image morphology clip and road boundary constraints, we fill the hole of road region reasonable in the detection, then complete and smooth it, and get the final detection result.The experiment under various kinds of scenes shows that the algorithm is not influenced by the illumination conditions, road texture diversity, road scenes and other factors. Meanwhile, the algorithm improves the detection rate and the incomplete problem of detected road region.