Crop edge detection based on stereo vision

Abstract

This paper proposes a robust detection method of uncut crop edges which is used for automated guidance of a combine harvester. The utilized stereo vision system allows for real-time depth perception of the environment. A three-dimensional elevation map of the terrain is constructed by the point cloud acquired in this way. The heights of crop and harvested areas are estimated using Expectation Maximization and segmented using of the clustering results. In a row-wise processing step, each scan line of heights is cross-correlated with a model function to compute possible candidate points located at the very crop edge. Using robust linear regression, a linear crop edge model is calculated, modeling the spatial distribution of the candidate points. An overall crop edge model is updated via exponentially weighted moving average.