Research On Obstacle Detection For Robot Under Comples Off-road Environments

Environment perception technique is the basis of autonomous navigation for unmanned vehicle, and obstacle detection is one of the key aspects of environment perception. When unmanned vehicle is moving in field environment, the ability to detect various complex obstacles is required. On the one hand, a wide variety of obstacles, like rocks, trees, puddles, rivers etc., are threatening the unmanned vehicle’s safety; on the other hand, vegetation and obstacles are mixed together in field environment, increasing the difficulty of detection.Obstacle detection in complex field environment is divided into two parts in this article. One is about the detection for different types of obstacles and complex terrain, and the other is about the obstacle detection in vegetation-covered environment. Then these two aspects are studied respectively.For the detection for different types of obstacles, the method based on64-line LIDAR is proposed in this article to detect protruding obstacles, deep groove, side slopes and waters. Considering of the potential danger in the area that sensors fail to detect, a method of classifying the unknown regions is also proposed. A method based on height information in the grid is adopted to detect protruding obstacles, and the suspension structure is optimized. A method based on the principle of LIDAR scanning is adopted to detect deep groove and side slope by analyzing the geometric feature of point cloud, and a representation of point cloud based on polar coordinate grid is also proposed. Water detection is realized in LIDAR intension image, taking advantage of the special reflective properties of waters. Experiments under field environment have verified the effectiveness of the algorithm. All the above mentioned functions have been realized and the detection rate is satisfying.For the obstacle detection in vegetation-covered environment, a detection method that integrates distribution features of three-dimensional point cloud and multi-spectral signature is proposed in this article. First, a multi-sensor system consisting of color camera, infrared camera and three-dimensional LIDAR is introduced, and a joint calibration method by camera and three-dimensional LIDAR as well as the integration process of three-dimensional point cloud data and image pixel data is described. Then, based on multi-spectrum data analysis and Normalized Difference Vegetation Index (NDVI), a new spectral signature of IR-color joint channel is proposed and used to classify vegetation and non-vegetation objects together with Gaussian Mixture Model. However, it was found in the experiment that the changes of lighting conditions and the interference of ground points had a significant impact on the results. Then by adding normalized light intensity of infrared light and weighted feature information, the detection effect was significantly improved. Experiments were carried out in several typical scenarios, and results showed the detection effect by this method was better than the one by NDVI.