Automatic detection of crop rows based on multi-ROIs

This paper presents a new automatic and robust crop rows detection algorithm for a vision-based agricultural machinery guidance system. The algorithm consists of five steps: gray-level transformation, binarization, candidate center points estimation, real center points confirmation and crop rows detection. During the procedure of candidate points estimation, a multi-ROIs is constructed, which integrates the features of multi-rows based on a geometry constraint that inter-row space is approximately equal into a single processing of optimization. This strategy ensures the center points of crop rows could be detected in a high weedy field or in a large barren patch of the field. For the fourth step, we used the clustering method to confirm the real center points indicating crop rows. And lastly, the linear regression method was employed to fit the crop rows. Three image data sets, taken from a wheat, corn and soybean under different natural and field conditions, were used to evaluate the detection rate, detection accuracy and processing time of the algorithm. Thorough an experimental comparison with the standard Hough transform (SHT), it demonstrates the proposed method outperforms SHT either in detection rate, detection accuracy or computation time. The algorithm requires about 61 ms to recognize crop rows for a 640 × 480 pixels image while the detection rate reaches 93%, detection accuracy reaches 0.0023°.