A Vision-based Road Detecting Algorithm For Outdoor Mobile Robot

In this article, some researches and analysises are carried to the experimental environment of outdoor mobile robot. And a road detecting system based on vision message is proposed here. It mainly includes: the road region detecting system based on vision message and the road'edge detecting and following system based on vision message. Massive actual experiments have been carried to confirm its validity and real-time ability.A new detection method for unstructured road based on robot's vision is proposed to improve the effectiveness of the road in complicated environment. First, OTSU is used to classify the road image. Then, the referenced region of road is adopted to judge the quality of the first classification. If the first classification is not satisfying, the OTSU will be used again to depart the road image to four classes. By combining the second classification result and the referenced region of road, the road and non-road model is built to determine the final result. This algorithm shows good self-adaptive ability and only needs little priori knowledge. It is also robust against noises, shadows and illumination variations and shows good real-time performance. It has been tested on real robot and performed well in real road environment.On the foundation of detecting road region, two road's edge detecting and following algorithms are proposed focusing on side camera. One is the method basing on the combination of Canny operator and HOUGH transformation. This method used the virtue of the edge which the Canny operator examined having single-pixel width and the virtue of its strong filtering ability. Then HOUGH transformation is used to fit the straight line. Another method is based on combination of Sobel operator and least squares algorithm. The gradient value and the gradient angle value which are computed by Sobel operator are used to forecast the edge points appearing in next frame. Then the least squares algorithm is used to fit these edge points. These two methods are all validated in massive experiments on real robot. The experimental results indicated these two algorithms all have better real-time ability and robusticity. It is suitable and feasible.