Research On Agricultural Environment Reasoning Based On Conditional Random Fields

The development of modern agriculture requires higher efficiency, better quality and lower operating costs. It is required to have higher level of automation and intelligence for agricultural machinery, so agricultural robot emerges. Before agricultural robot operates autonomously in agricultural environment, it needs to understand agricultural environment and obtain the scene knowledge. By using visual sensor to obtain agricultural environment information, this paper presents an image segmentation and classification algorithm, which uses super-pixel as the image processing unit, and integrates image spatial context information by conditional random fields. The main contents and conclusions are as follows:To deal with the uncertainty and ambiguity of agricultural environment, image spatial context information is integrated by conditional random fields to improve the accuracy of segmentation and classification. To deal with the problem that using a single pixel as image processing unit is susceptible to noise and reduces the efficiency of algorithm, super-pixel as image processing unit is presented in this paper, and neighborhood super-pixel is defined.A pear scene segmentation algorithm is introduced. Firstly, a scene image labeled on hand is segmented into super-pixels, and then the feature vectors and labels of super-pixels are integrated as samples into class databases. Secondly, the model by using conditional random fields to integrate image spatial context information of unlabeled super-pixels is constructed. Finally, after getting model parameters the labels of unlabeled super-pixels are inferred through the maximum posterior criterion. Compared with the nearest neighbor algorithm, the pear scene segmentation algorithm is more adaptable for opera natural environment with uneven distribution of illumination, random distribution of weeds and content mutation.A terrain reasoning algorithm based on conditional random fields is built. Self-developed agricultural mobile robot is as the platform, using a binocular camera to get the label information online, and the cluster centers are built and updated by the incremental integration strategy. The result of confirmatory experiment demonstrates that the proposed algorithm is adaptive under the circumstance of change in light and uneven light distribution.