Extraction of visual navigation directrix for harvesting operation in short-stalked and close-planting jujube orchard

Abstract: An algorithm based on image processing technology was proposed for generating navigation directrix in complex circumstance of short-stalked and close-planting jujube orchard. The hatching method was used to analyze the distribution of target pixels and then proper processing image was obtained. Tree trunk and background were segmented based on B component of image which was binarized by threshold of each scan line. By analyzing the distribution of crossing points between tree truck and ground, the chief processing section, or the region of interest (ROI), was defined and chosen from two-fifths of image area below. Then in the light of trunk longitudinal gray scale distribution, tree trunks location was extracted with gray scale vertical projection method and morphology principle by setting a superficial window to dynamically scan ROI. In order to describe line trend in short-stalked and close-planting jujube orchard, trend lines were introduced, which included the right one and the left one in terms of ridge and furrow. To describe linear distribution trait of ridge and furrow, crossing points between tree and ground were obtained by comparing the shortest distance of candidate point to its corresponding trend line with a man-made threshold value. Afterwards, those selected points were classified separately into 2 clusters in terms of their distribution region which was located at right or left part of image longitudinal symmetry axis, with the point set available to fit the border line of ridge separately. The least square method was used for detecting the right and left border lines, and navigation directrix was generated by extracting the center points between 2 border lines. The method of extracting navigation path was searched by analyzing the condition of short-stalked and close-planting jujube orchard under harvesting operation, and the accuracy and reliability of the algorithm were analyzed and evaluated under a variety of environmental conditions. The study of the algorithm was still in the simulation stage and the specific navigation effect of the algorithm was related to the actual navigation operations, so algorithm performance could not fully represent the actual navigation applications. In order to measure the algorithm reliability and real-time parameters, 5 different scene conditions, which included 4 single factor working conditions and one multiple factor working condition, were tested. The experimental results showed that the algorithm could generate navigation directrix accurately and showed a good noise robustness. Under 4 single factor working conditions, the accuracy was more than 81.3%, and the average processing time consumed was less than 11.9 s to each frame image; by video detection, the accuracy was more than 83.4%, and the average processing time consumed was less than 1.7 s to each frame image. Under multiple factor working condition, the accuracy only reached 45% and the average processing time consumed was 9.4 s; by video detection, the accuracy only reached 42.3% and the average processing time consumed was 1.0 s. Therefore, for subsequent tasks, work should be done to improve the real time performance and practicality of the algorithm for various surroundings, so as to enhance the sensitivity of the algorithm under the multi-mode condition. The novelty of this paper is to propose several new methods to solve practical problems. The research provides a reference for autonomous navigation of robot in short-stalked and close-planting jujube orchard.