Search Results

A machine vision system was developed to identify and locate harvestable spears of asparagus. An image acquisition vehicle was fabricated to videotape portions of asparagus rows from a commercial production field. Images were acquired using a monochrome CCD camera. The detection of reflectance properties of asparagus was enhanced by using optical bandpass filters for near-infrared radiation. Videotaped segments acquired in the field were analyzed. Image processing techniques based on geometrical characteristics of asparagus spears were used to identify and locate harvestable spears in the images. Harvestable spears measured in the field were compared to those found by machine vision. The vision system correctly identified from 86 to 97% of the harvestable spears in six 15 m row segments analyzed. The uncertainty in the location of spears was within a 2.97 by 5.39 cm window with 95% confidence.

A heuristic line detection algorithm is described for computing guidance information from row crop images. The technique processes binary images representing crop rows against a soil background. Points along the centers of crop rows are enhanced using a modified run-length encoding procedure. The properties of lines in images can be improved by filtering based on characteristics of the object run-length. A clustering algorithm was used to aggregate pixels that fall on the same crop row. The technique was compared with the Hough transform, a common line detection technique in image processing. Both procedures accurately represented lines measured manually in a set of images representing a range of expected field conditions.

Image processing techniques were investigated for developing a guidance signal for a tractor operating on agricultural row crops. The guidance signal was computed from thresholded images segmented by a Bayes classifier. The distribution of crop canopy and soil background pixels in an image was approximated with a bimodal Gaussian distribution function. The parameters of the distribution were estimated by regression to systematically subsampled images. Run-length encoding was used to locate center points of row crop canopy blobs in the thresholded images. A heuristic line detection algorithm was used to determine the parameters defining crop row location on the image plane. Row parameters were used to compute a tractor guidance signal. Results are presented on the performance of the individual components of the algorithm.