Research On Strawberry Recognition And Localization Based On Deep Learning And Binocular Vision

Today,with the continuous development of agricultural mechanization,making full use of agricultural robots to replace traditional manual operations can not only greatly reduce the labor intensity and labor cost,but also improve the labor efficiency.The picking robot is an example of the application of artificial intelligence technology in the agricultural field.It consists of a power drive system,a control system,an information collection and perception system,and an instruction execution system.It can realize the perception of the surrounding environment and the required information to a certain extent.Functions such as independent decision-making.In a complex environment,the existing picking robot vision system has certain difficulties in identifying and locating fruits.Aiming at these problems,this paper selects strawberries with higher economic returns as the research object,and conducts identification and positioning research.(1)A binocular stereo vision system is established,and related parameters of the binocular camera are obtained through camera calibration,and then the images collected by the left and right cameras are stereo corrected according to the obtained parameters.(2)Under the Tensor Flow framework,Inception v2 was selected as the feature extraction network,and a strawberry data set was made,and Faster R-CNN was used to detect strawberries.The final average recognition accuracy is 85.32%,the average time for identifying each image is 0.16 s,the detection accuracy and speed basically meet the needs,and the recognition results can provide data for positioning work.(3)Use the region-based stereo matching algorithm BM to complete the stereo matching of the strawberry picking area,and conduct the ranging accuracy experiment.The error analysis shows that the system has an average error of less than 3% in the range of 200mm-2000 mm.(4)Three-dimensional positioning experiment is carried out on the strawberry according to the position given by the target detection.The error of the positioning obtained through error analysis is less than 5%,which can basically meet the three-dimensional positioning requirements of the strawberry picking robot.