Method for detecting crop rows based on binocular vision with Census transformation

Machine vision is an important method for the navigation of agricultural machinery. The crop row detection is a key aspect of the navigation based on machine vision for pathway determination. The binocular vision based technique can locate the spatial position of crop rows, which is more effective for crop field with high weeds pressures than monocular vision based technique. The feature point detection and stereo matching are essential aspects of binocular vision algorithm, which will affect the accuracy and efficiency of crop row detection. To enhance the accuracy and efficiency of crop row detection for complex crop field, a new crop row detection method based on Census transform was presented in this paper. The presented method consisted of 3 modules that were image preprocessing, feature point detection and crop row detection. The image preprocessing module was composed of grayscale transformation and feature point detection. To separate crop rows from backgrounds, an improved function of excess green minus excess red (ExG ? ExR) was used to transform RGB (red, green, blue) color image to grayscale image, in which only the ExG features were transformed by the classical ExG ? ExR function. The improved ExG ? ExR function was compared with the ExG and the classical ExG ? ExR method. The comparative results showed that the improved ExG ? ExR function was more effective to suppress the background noise. The smallest univalue segment assimilating nucleus (SUSAN) detector was used to detect corner points, which could describe the contour of crop rows. The module of feature point detection consisted of stereo matching and three-dimensional (3D) coordinate point calculation. An accurate stereo matching algorithm based on Census transformation was used to calculate the disparity of SUSAN corner point. The Census transformation of the SUSAN corner point was the primitive, and the sum of absolute difference (SAD) function was the similar metric. To reduce the computing burden and improve the accuracy, the sizes of Census mask and SAD mask were both 5×5 pixel. As the binocular camera (BB2-08S2C-38) used in this paper was assembled with 2 parallel monocular cameras, the 3D coordinates of SUSAN corner points were calculated based on their disparities. If the coordinates in the width and height axes of an SUSAN corner point were within the range of width and height thresholds, the point would be extracted to be a feature point of crop rows. As the crop rows were always parallel according to agronomic arrangement, the amount of crop rows in image could be estimated based on the frequency histogram of width distribution. The distance between adjacent crop rows was assigned to the interval, and the range of coordinates in width axis of feature points was assigned to the number of groups in the histogram. The feature points on each crop row were distributed with the shape of ellipse. After obtaining the amount of crop rows, feature points were used to locate the long axis of their distribution based on the principle component analysis (PCA) method. The long axis of the feature points was used to fit the centerline of the corresponding crop row. An SAD based crop row detection algorithm was set as a comparative algorithm. Videos of cotton field that were without infestations, with shadows, with weeds and in turnrows were used to test the 2 algorithms. Results showed that, the proposed algorithm was more robust to the changes of conditions and consumed a little more time; in the situations without turnrows, the accuracy of crop row detection was no less than 92.58%; for the deviation angle of detected centerlines of crop rows, the absolute mean value and standard deviation value were no more than 1.166 and 2.628° respectively; for the processing time, the mean value and standard deviation value were no more than 0.292 and 0.025 s. The accuracy and efficiency of presented method can satisfy the requirement of field operations.