Abstract: At present, the problem of environmental pollution and ecological deterioration raised by agricultural production is increasingly serious, so the development of the agricultural robot guided by machine vision is widely considered to be necessary which has the potential of precision farming and being friendly with natural environment. In this paper, an algorithm of road recognition is introduced, which is a key technology for the visual guidance in the non-structured agriculture environment. At first, the suitable segmentation feature of the color picture of natural field environments is discussed. Based on the road knowledge and the feedback idea, then a new edge following method is developed to recognize the road. Finally, compared with the conventional method of threshold segmentation, the result of the treatment of actual field pictures shows that this algorithm is more effective for the road recognition.