Vision-based Lawn Boundary Recognition for Mowing Robot

Abstract

We present a novel method for mowing robot's lawn boundary recognition based on Gabor filters and support vector machine (SVM). Robust texture features of images are extracted and concatenated using Gabor filters. The principle components analysis (PCA) approach is then used to reduce the dimensionality of Gabor features. Based on the compressed features, SVM model is trained and used to perform the grass texture classification task. The boundary of lawn is then recognized according to the ratio of grass area of the image. To demonstrate the effectiveness and robustness of our proposed method, a dataset is created with about 1500 images of different lawn scenes. Result shows that a classification accuracy of 96.7% can be reached when SVM is used. Experiments of the lawn boundary recognition have also been conducted on the mowing robot under different lighting conditions. The recognition rate tested is 98.3%, which proves the efficiency and superiority of our proposed method.