Research On Navigation Path Recognition Method Of Wheat Harvester Based On Semantic Segmentation

Accurately identifying the boundary between the harvested and unharvested areas of wheat makes the wheat combine autonomously navigate along the crop boundary,which greatly improves the field operation quality and efficiency,and reduces the labor intensity of the wheat combine manipulator.The existing navigation path recognition algorithm of wheat harvester based on machine vision is difficult to adapt to the complex and changeable wheat harvest environment,and the model requires high computational power,which is difficult to run quickly in the vehicle terminal.Therefore,this study proposes a navigation path recognition algorithm for wheat harvester based on deep learning semantic segmentation,which can realize accurate and rapid recognition of target areas from complex environments.The main research work is as follows :(1)The data sets of six typical wheat harvest area scenes were constructed.In order to make the research on the navigation path identification of wheat harvesters more universal,the images of wheat harvest areas in six typical environments,including strong light,weak light,inverse light,fair light,shadow and plot edge,were collected.After the images were enhanced,the Labelme software was used to label the collected images,and the data set in VOC format was made,including 1396 training sets,478 verification sets and 478 test sets.Prepare for subsequent semantic segmentation model training.(2)A dataset of 6 typical wheat harvesting area scenarios is constructed.In order to make the research on the recognition of the navigation path of the wheat harvester more universal,the images of the wheat harvesting area under 6 typical environments,including strong light,weak light,backlight,smooth light,shadow,and plot edge,were collected,and the images were analyzed.After the enhancement processing,Labelme software was used to label the collected images,and they were made into VOC format data sets,including 1396 training sets,478 validation sets,and 478 test sets.Prepare for the subsequent training of the semantic segmentation model.(3)Based on Mobilenetv3,the basic semantic segmentation model Deeplabv3+ is designed and implemented with secondary optimization.Based on the research of(2),Mobilenet V2 is replaced by Mobilenet V3-Large as the feature extraction network,and Leaky＿Re LU is replaced by Re LU as the activation function,and its structure is trimmed;at the same time,the hole convolution is replaced by a depthwise separable convolution.The spatial hole convolution pooling pyramid module in the network decoder is optimized.The comparative analysis of the model training shows that the segmentation accuracy and the intersection ratio of the wheat to be harvested area are98.04% and 95.02%,respectively,which are 0.35% and 0.38% higher than those before optimization;the video image processing frame rate FPS is 7.5 frames/ s,1.2 times before optimization.(4)Accurate and rapid identification of navigation paths in six typical wheat harvesting scenarios is realized.First,use the semantic segmentation model trained in(3)to segment 6 kinds of wheat harvesting scene images,and obtain their segmentation mask map;secondly,extract the key of the navigation path of the wheat harvester in the segmentation mask map by the horizontal scanning method information points,and then use the multi-segment cubic B-spline curve fitting algorithm to realize the identification of its navigation path.After testing and analysis,it is known that the maximum pixel error and average pixel error of navigation path recognition are 11.7pixels and 7.4 pixels respectively,the maximum distance error is 0.063 m,and the average distance error is 0.037 m.The time is 0.15 s.To sum up,the method for identifying the navigation path of wheat harvester proposed in this paper can better adapt to various typical typical wheat harvesting scenarios,and has the advantages of high accuracy,good stability,lightweight model and fast running speed.,which can better meet the work requirements of combined harvesting real-time navigation,and can provide theoretical basis and technical support for improving the autonomous navigation capability of wheat harvesters.