| The harvesting operation of a wheat combine harvester is an important link in wheat mechanized production.Precise segmentation of wheat harvested areas and non harvested areas and planning of navigation paths can greatly reduce the intensity of manual operations and improve the efficiency of wheat harvesting operations.Due to the complex field environment of wheat harvesting,the current wheat harvester navigation algorithms based on machine vision are difficult to accurately segment boundaries,and they require large amounts of memory and high computational power requirements.There are certain technical difficulties in applying embedded computing platforms to combine harvesters.This study proposes a method for wheat harvester operation boundary segmentation and navigation based on an improved Ghost-UNet semantic segmentation model.It is possible to accurately segment the harvested and non harvested areas of wheat in complex field environments,and establish navigation paths.Moreover,the model occupies a small amount of memory,enabling the application of an on-board embedded computing system on a wheat combine harvester.The research content of this article includes the following aspects:(1)Based on the analysis of image segmentation methods,the segmentation effects of different segmentation methods on wheat harvest images were compared and analyzed.Traditional segmentation methods are susceptible to conditions such as color,light,and noise,and have poor segmentation effects on harvested wheat stubble areas and non harvested plant areas.Using a semantic segmentation model based on deep learning has strong robustness and generalization capabilities for color,light,noise,etc.,and can accurately and efficiently segment the harvested wheat stubble area and the non harvested plant area in the wheat harvest image.This study compares and analyzes different semantic segmentation models,and finally selects a UNet model that has better segmentation effects on edge details to segment the wheat harvest image.(2)Collect wheat harvest images under different harvest scenarios and construct a dataset.After comparing and analyzing lightweight models such as Squeeze Net,Mobile Net series,Xception,and Ghost Net,we finally selected Ghost Net,which has the best model optimization,fast computation speed,and high accuracy,as the lightweight backbone model;Introducing the attention mechanism of CBAM and the idea of transfer learning,a Ghost-UNet semantic segmentation model is proposed to segment the wheat harvest boundary.Finally,various indicators such as MIOU,Recall,PA,and MPA of the Ghost-UNet model for wheat harvest segmentation were obtained to evaluate the segmentation performance of the model.(3)On the basis of wheat harvest boundary segmentation images,black and white binary and morphological processing are performed on them to facilitate subsequent boundary key point extraction.After that,horizontal scanning method is used to obtain boundary key information points at equal intervals,ensuring accuracy while reducing the amount of computation.Then,the wheat harvest boundary line is fitted using the minimum quadratic straight line,and finally,image fusion is performed to establish the wheat harvester navigation path in the original image.(4)Establish a wheat harvester navigation path recognition system.Select Jetson Xavier NX as the embedded computing platform,complete the environment and algorithm configuration,and carry out field test verification on a wheat harvester;Field experiments were conducted in different environments,different lighting,and different backgrounds.The results showed that the navigation path recognition system established in this study can recognize wheat harvest boundaries in different scenarios by more than 94%,indicating that the crop boundary line recognition method proposed in this paper can effectively and accurately segment wheat harvest boundaries in various complex wheat harvest scenarios,and thereby establish wheat harvester assisted navigation paths. |
PDF Full Text Request