Approach Of Visual Identification And Path Planning For Apple-picking-robot

To identify the apple fruits in tree canopies effectively and precisely, to detect theobstacles occurred between lines in an orchard at the walking forward direction of robotbefore being picking, and to support methodologically the path planning technology whileapple-picking-robot being operated, the study on visual identification and path planning hassignificant effect and potential practical value on improving the reliability and pickingefficiency of the apple-picking-robot, and on promoting the intelligently picking level ofagricultural equipment.To solve the key points of visual identification and path planning for apple-picking-robot,this dissertation focused on apple in canopies recognition under a non-structured naturalenvironment, on uncertain obstacle detection at the walking forward direction between lines,on completely coverage path planning with non-learning mechanism bio-inspired neuralnetwork-based and on path planning with Q-learning mechanism. The proposed approacheswere implemented and verified using Matlab2010a in Window7. In this study, the maincontents and original results were described as follows.(1) A novel approach to detect natural images in an unstructured canopies with fuzzycolor judgment and L*a*b*color space was proposed. The experiments on apple detection incanopies showed that if apples emerged red over10%even with some shades, they could beeffectively and totally detected; and the experiments on uncertain obstacles detection ingrass-planted orchard also revealed that if most the background appeared green even withsome branches and trunks, the obstacles could be effectively and totally detected within lines;the grey levels of the background could be close to zero, and the object could be highlightedafter completion of a*component conversion into gray image in L*a*b*.(2) A novel thresholding method based on maximum entropy theory was proposed tosegment the apple gray-scale image after the completion of gray image conversion. Theexperiments showed that the running time of segmentation on apple gray image with Renyientropy theory was about16milliseconds, and of segmentation on obstacle image with fuzzy2D entropy with optimized genetic algorithm about34-36milliseconds, which could segmentthe object from background effectively and clearly, be regardless of the size of image andnumber of objects detected, and satisfy the real-time requirements of picking robot. (3) A new apple shape features extraction method with the centroid and roundness ofconnected regions was proposed. The experiments verified that these proposed approachescould effectively recognize apple from173images in tree canopies under a complexenvironment, other7images failed for the reason that apples emerged red less than10%withserver light exposure, and the recognition rate was96.1%, which could provide a new trial forthe fruit detection and recognition. A new filtering method with mathematical morphologywas proposed to eliminate the sticky connected regions of branches and trunks using the ratiobetween perimeter and area of the connected regions over0.375, and the obstacle detectionratio was96.9%in the experiments of98images of grass-planted orchard.(4) A novel global path planning of zigzag complete coverage approach based onbio-inspired neural network was proposed according to the characteristic of parallel plantinglines. The experiments under static known and dynamic unknown environments verified thatthe activity of neurons located at unpicked area could all along attract the robot until the robotreached at the appointed position after the completion of construction of discretizedtopological map using non-holonomic constraint kinematics. The running time for pathproducing was about50milliseconds. The whole planning process need not learn, and hashigh universality, lower running cost, less turn control, and could be suitable for path planningor target tracing under complex environments.(5) A novel apple-picking-robot with Q-learning mechanism was designed to ensurerobot moving or working safely including the design of avoidance of obstacles, and pathexploring policies and learning mechanisms, while the operating environment was changed sorandomly and abruptly as that the designer could not understand the whole environmententirely. The experiments under static known and completely unknown environments verifiedthat the new robot could find the prior and safe path from the starting point to appointed point,which could guide the new robot move towards after the completion of10minutes’ learningwhile the pre-estimated obstacles number including static and random obstacles was no morethan10. The Q-learning method was regardless of environment, and suitable for path ortrajectory planning and complex control under uncertain or unknown environments, and couldpromote the adaptability of robot to process the abrupt incidents. Like other learningmethodologies, the only disadvantage was that the learning process was a littletime-consuming for the reason of high-dimension data disaster.