Research On Farmland Navigation Technology Based On Scene Understanding

With the continuous development of intelligent agricultural machinery technology,navigation technology,as an important part,has become the hot research object of many experts,in which the navigation line extraction is the key to realize precision agriculture and automatic navigation.In order to solve the problem that most farmland scenes are unstructured,a novel navigation line extraction algorithm based on semantic segmentation results was proposed in this paper.In this study,we used the Kinect 2.0 depth RGB-D camera developed by Microsoft,and take the wheat field as the research object,proposed a method to realize scene recognition at the semantic segmentation level by combining color information with depth information to construct a kind of HHG data,and extracted navigation line based on this result.The main research contents of this paper including:(1)Data acquisition and preprocessing.The training speed of the subsequent neural network is accelerated by filtering,cutting,graying,marking,enhancement,fusion and other operations on the collected color data and depth data.(2)Semantic segmentation.We collected three kinds of data in the wheat field scene at the same time,including color,depth and 3D point cloud data,each data type has 707 data.The semantic segmentation model trains 508 image data randomly selected from all the collected images,and the rest is for the test set.The FCN is used to segment wheat,ground and background in the fused image data,and the MIOU is obtained 95.4%.(3)Navigation line extraction.According to the characteristic that the result of semantic segmentation is distinguished by different colors,the region of interest(ROI)is extracted by color threshold,and the boundary line is found by line-by-line scanning in the part of ROI,and then the navigation line is calculated.The average angle deviation of the final pilot detection is 0.08°,and the average distance deviation is 0.047 m.In this paper,we mainly used the HHG data which fused the color information and depth information to complete navigation line extraction,and added the other two comparison methods.The experimental results show that,compared with the comparison method,the method proposed in this paper can better perform semantic segmentation and navigation line extraction in the same wheat field scene,and the method is robust to weed pressure.