Maize stubble row recognition and guidance line detection based on machine vision in natural illumination

Abstract. In North China Plain where annual maize-wheat rotation is predominantly managed, vision-based automatic guidance can be used for direct seeding of wheat in no-till field, to avoid the standing maize stubbles and the subsequent planter blocking during seeding operation. This study aimed to focus in the recognition of maize stubble row and the detection of guidance line. Static images of maize stubbles standing in the field were acquired in natural illumination after the combine harvest of maize, separately on cloudy days and sunny days. Sample images extracted from the collected statistic images experienced a color feature analysis in RGB and HSI color spaces. Due to the large overlap of the standing stubble and background in previous analysis, a region of interest (ROI) was selected to reduce the difficulties in subsequent segmentation. A color index ‘0.44R+0.55G-0.5B‘, which is obtained by optimizing the linear combination of RGB components basing on genetic algorithm, was used to transform the ROI into monochrome. The threshold for ROI binarization was determined after multiple iterations, meanwhile an evaluation criteria was proposed to eliminate pixels that were judged as noise from the binary ROI. Afterwards, least square method, the commonly used algorithm for linear fitting, was adopted to fit the guidance line considering its superior processing efficiency. And a noise elimination method was developed to reduce the negative effects of noisy point on baseline fitting. In the end, the proposed maize stubble baseline detection algorithm was tested using 30 randomly selected field images, 73.1% of the total resulted in effective detection.