The Research Of Road Scene Understanding Technology

With the rapidly development of technology, the role of vehicle intelligent system is becoming more and more important in computer vision. Being an important part of the vehicle intelligent system, the road scene understanding will naturally become the hotspot of the research field. Scene understanding is a deeper level of the object recognition field, which is based on the image analysis. There has been a lot of exploration about the scene understanding technology and many inspiring achievements have been made. The current scene understanding technology is mainly used in intelligent robots, intelligent vehicles and other equipment, and it is usually based on the video stream or panoramic view, while the study about the single road image is not enough. How to conduct scene understanding based on a single static image is still a challenge for the instable shape and inconsistent color of road. It is also easily affected by light and noise factors, which make it more difficult to understand the scene. This paper discusses the key technology of road scene understanding which contains the segmentation of image, vanishing point detection, road boundary detection and depth estimation. In the area of image segmentation, we present image segmentation method based on threshold, edge and cluster analysis. In the field of vanishing point detection, we introduce the vanishing point detection method on the basis of texture feature, and propose the improvement scheme which can detect the far vanishing point correctly; In road boundary recognition, we describe the road detection method based on feature and model, and propose a method to detect road boundary based on road edge and vanishing point; In the field of depth estimation, we introduce the relevant knowledge about it and describe the depth estimation method based on image segmentation results.In order to implement the segmentation and understanding of road image, we first use GMM(Gaussian Mixture Model) method for image classification, and propose a scheme which can automatically obtain the cluster center. On the basis of image classification, we use the edge image as a constraint to perform region growing, and we get the sky and road region. And then we use the texture orientation histogram to limit the vanishing point candidate area based on road region and conduct far vanishing point detection based on texture feature. After we get the location of vanishing point, we use hyperbola curve to fit the road boundary based on road edge image and vanishing point location. In the end, we conduct depth estimation based on road boundary and region extraction results to get the depth image consisting of sky, road and background.The experimental results show that the proposed algorithm is applicable to a variety of road scenes. It can also extract the road area and sky area correctly, get the accurate location of the vanishing point, extract the boundary curve model correctly, which has good robustness, and it provides the depth estimation method based on the road extending trend.