Plant recognition and localization for intra-row mechanical weeding device based on machine vision

Abstract: Intra-row mechanical weeding, as a non-chemical weed control technology, reduces the application of chemical herbicides and is beneficial to the environment protection and sustainable development for agriculture as well. Most crops are cultivated in rows with a defined sowing or transplanting pattern, i.e. with a constant spacing distance. This is an important feature that can be used for plant recognition and localization. The goal of this study presented herein is to propose a recognition and localization approach, taking advantage of the knowledge of the sowing or transplanting pattern, to avoid crop automatically and enter into the intra-row area for intelligent intra-row mechanical weeding device. The RGB imaged plants were distinguished from soil by analyzing the excessive green (2G-R-B) vegetation index image. The Ostu algorithm method was employed to transform a gray image to a binary image. And then the binary image was dilated and eroded three times repeatedly to remove isolated pixels in binary images or to remove noise for subsequent analysis. The standard deviation of longitudinal histogram was used as the scanning line to get the crop row area information in a binary image. The next step was to sum up all pixels of the crop row area per column, thus forming a signal with a frequency that corresponds to the average crop distance. The target regions and center points were obtained by analyzing the lateral histogram with the horizontal scan line. The most probable crop regions were filtered from all the target regions using a sinusoid which was fitted lateral histogram based on the distance between crops. The phasing of the sinusoid was given by least square fit for all the center points. After fusing the center of crop row and the centroid of green plants in binary image, the plants localization were obtained through searching the closest fusion result to the sinusoid peeks. Test results showed that, the method was sufficient in plants recognition and localization for intra-row mechanical weeding under different weather and field conditions. The accurate identification rate was 95.8% with the absolute error of 4.2 pixels in the x-direction and 1.4 pixels in the y-direction for cotton seedlings. An identification rate of 100% with the absolute error of 6.8 pixels in the x-direction and 15.3 pixels in the y-direction was achieved for lettuce seedlings. The position of the crop was correctly determined for 100% of all the images. The positioning error for lettuce and cotton seedlings was 17.6 pixels and 5.0 pixels, respectively. Main factors that influence the performance of the recognition and localization are weed pressure and the plant growth conditions. This study provides the basics for mechanical weed control devices to seedling avoidance and automatic weed control.