Efficient visual obstacle avoidance for robotic mower

We present a novel method for robotic mower's lawn obstacle avoidance as well as motion control, which is based on the Gabor texture classification and drivable region search methods. In our approach, a camera is applied to obtain real-time image streams of lawn scenes, Gabor filters are then applied for extracting robust texture features. Based on the compressed features, a SVM model is trained and used to perform the grass texture classification task. The maximum connected drivable region is computed via a breadth-first search method (BFS) according to the classification results. By restricting the BFS method in a polygon window area, we accelerate the region search step by 64% compared with searching the whole image. The whole program is built on the Robot Operating System (ROS) for the purpose of expansion. Experiments have been performed on the robotic mower under complex scenes, including various obstacles, seasons and lighting conditions. The obstacle avoidance rate tested is 97.7% in a 75 m∗40 m area, which proves the efficiency and superiority of our proposed method.