Research On Recognition Of Navigation Scenarios For Agricultural Robot Based On Conditional Random Field

As we all know, the mobile agricultural robot always works in different complex environments, if we wish it completes the work independently, it must be equipped with certain perceptions to learn the environmental terrain independently. The commonly methods on the terrain understanding are the binocular vision and radar, etc. They can estimate where the different scenarios regions are. Due to the unstructured scenes of randomness, diversity, complexity and robot motility, this aggravated the recognition of the unstructured environmental regional terrain accurately. So, how to use the sensors to detect the operation environment similar to human environment cognitive ability is especially important, and which would make agricultural robot more accurate understanding of its working environment. Therefore, this article based on conditional random field (CRF), one of the undirected probability graph model, and with the introduction of super-pixel image segmentation, combined both of them finished the research, main contents and conclusions are as follows:In order to improve the area of recognition effect and identification efficiency, in scenario regions identification process, we firstly use graph-based super-pixel segment work to the scene. Super-pixels corresponding to local connected brightness similar pixel sets, by using this segment method improved the scene recognition efficiency, reduced the time cost. It would be helpful to subsequent agricultural robot autonomous navigation research.According to the unstructured environment uncertainty and the scene pictures always contain randomness and complexity, we did not only consider the requirements of the recognition efficiency, but also gave consideration to the scene regional identification recognition accuracy. Therefore, in this paper, we put CRF into scene segment to deal with the environment scenery of unstructured characteristics and randomness, and the scenes in each part of the relevance scenarios, etc. Through the fusion of the two parts, the scene recognition algorithm has some adaptability, which accurate understanding of the unstructured indoor scenes and natural environment terrain category.Finally, we put our recognition algorithm into the agricultural robot platform which was designed by ourselves, and experimental results showed that the agricultural robot effectively able to complete the recognition of the scenarios regions with more of availability and reliability.In this paper, navigation scenarios regions recognition of agricultural robot researched based on conditional random field, which provided the necessary theoretical preparation and one of the effective real-time scene recognition methods to realize agricultural robot autonomous navigation.