Visual Terrain Classification for Outdoor Mobile Robots

In this thesis we present a comparison of multiple approaches to visual terrain classification for outdoor mobile robots based on local features. For this purpose, we put a camera on a mobile robot and use it to capture images which are then analyzed to recognize the terrains present in these images. There are two sets of approaches that we use to classify terrains. The first is based on greyscale images and the second one is based on color images. For greyscale images, we use two different robot platforms for two different scenarios. The first robot platform is a wheeled outdoor robot. The second platform is a flying robot. For terrain classification, we modify and test three approaches called SURF, Daisy and Contrast Context Histogram, which are traditionally not used for texture classification. We compare these with more traditional texture classification approaches, such as Local Binary Patterns (LBP), Local Ternary Patterns (LTP) and a newer extension Local Adaptive Ternary Patterns (LATP). The image is divided into a grid and local features are calculated on the cells of this grid. These features are then used to train a classifier that can differentiate between different terrain classes. Images of different terrain types are captured using a single camera mounted on a mobile outdoor robot. We drove our robot under different weather and ground conditions and captured data of different terrain types for our experiments. We did not filter out blurred images which occur due to robot motion and other artifacts caused by rain, etc. We used a Random Forest classifier for classification and cross-validation for the verification of our results. It is shown that SURF features perform better than other descriptors for smaller cell sizes and LTP works best for larger grid cell sizes. The results show that these approaches work well for terrain classification in a fast moving mobile robot, despite image blur and other artifacts induced due to variant weather conditions. Furthermore we investigate the effectiveness of local image features for visual terrain classification for outdoor flying robots. A quadrocopter fitted with a single camera is flown over different terrains to take images of the ground below. Six different terrain types are considered in this approach. The images captured have artifacts like blur and scale variations. It is shown that SURF features also perform better here than other descriptors for smaller grid cell sizes and LTP performs better for larger cell sizes. We also test color image based terrain classification. For this purpose, we use two different types of camera mounted on our wheeled outdoor robot and capture five different terrain types traversed by the robot. We use two different image descriptors that can work on color images. The first descriptor is the co-occurrence matrix and the second descriptor is the SURF descriptor. Each of these descriptors is applied to the color