Obstacle detection is an important aspect of environment perception of mobile robot. In traditional obstacle detecting methods, parameters should be adjusted manually so as to accommodate variable outdoor environment. The introduction of machine learning can reduce the manual participation and improve intelligence degree of mobile robot, and can deal with the high-dimensional input information better as well.Traditional supervised learning methods only use labeled instances to study. However, due to the complex working environment, obstacles are unpredictable. In order to detect obstacles as accurately as possible, establishing a training database with humorous labeled instances is necessary, which is time-consuming and labor-intensive.Considering above problems, Semi-supervised Learning methods are introduced, e.g. Self-Training and Co-Training. We improve the learning performance by establishing an initial classifier with a small number of labeled instances, and then continually updating it with a large number of unlabeled instances.In this paper, a lot of experiments are carried out on environment-perceptive images. Sub-block strategy is used to extract the color and texture feature to obtain a training database with a small number of labeled instances and a large number of unlabeled instances. Based on the kNN classifier, Self-Training and Co-Training algorithms are introduced. Experimental results show that, by the introduction of Semi-supervised Learning, it can achieve a higher correct rate on both obstacle detecting and classification with only a small number of labeled instances. |