| Autonomous obstacle avoidance technology for robots is currently a research hotspot.In the research of orchard robots,the key technologies to realize orchard automation are to realize independent spraying,fertilization and ploughing.When performing such operations,the orchard robot needs to work close to obstacles and realize autonomous navigation.In order to improve the accuracy of obstacle avoidance,this thesis proposes to classify obstacles first,and then to study obstacle avoidance strategies for different types of obstacles.Firstly,in view of the complex environment of orchards and the variety of obstacles,this paper first proposes a method of obstacle identification by Convolutional Neural Networks(CNN).However,the image information collected by the camera is not high in quality due to environmental and other factors.Therefore,the image is preprocessed before recognition,including image filtering,image enhancement and edge detection.After preprocessing,Send it to the CNN trained in advance for identification.The algorithm can identify tree trunks,and its recognition efficiency reaches 73.8%.Secondly,according to the types of obstacles identified,the study of obstacle avoidance strategy is carried out.For roundness-type obstacles,the diameter and distance of the trunk are collected,and a circular obstacle model is drawn according to the parameters,so that the orchard robot rotates around the trunk after traveling to a certain distance,thereby avoiding obstacles.For irregular type obstacles,the orchard robot first draws the outline of the obstacles under the parameters according to the collected obstacle information to complete obstacle avoidance,and continues to obtain new obstacle parameters in the process,and then updates the outline parameters of the obstacles to complete obstacle avoidance.Through simulation and orchard test,this method is closer to obstacles than that of the traditional obstacle avoidance method,and is suitable for machinery operation in the orchard. |