Research On Citrus Target Recognition And Localization Method Based On Machine Vision

China’s citrus cultivation scale and productivity are in the forefront of the world,and in such a huge citrus production operation,citrus picking accounts for a large proportion of the operating labor.Nowadays,in the face of high labor costs,low work efficiency,labor shortage and other problems,the use of picking robots with a high level of intelligence to liberate labor is an inevitable choice.Therefore,studying a vision system that can accurately identify and locate citrus targets is of great value for automated harvesting.The purpose of this paper is to provide the necessary visual information for the production operation of citrus picking robots.Therefore,taking citrus fruits as the research object,the identification and localization of citrus targets in complex orchard environments are studied,and the research content mainly includes the following aspects:(1)The imaging model of the camera is analyzed,the principle of Zhang Zhengyou’s calibration method is introduced,and a self-calibration method is proposed to confirm the camera parameters by rectangular image,and the extinction point determined by the third set of parallel lines orthogonal to the two sets of parallel lines of the rectangle is solved based on the nature of the vanishing point.Then,the rotation matrix is further obtained through the geometric relationship of the three extinction points,and then the parameters and translation vectors in the camera are calculated according to the constraint relationship between the rotation matrix and the homology matrix of the camera.The method is easy to operate and simple to implement,simulation experiments show that the method has high robustness,while real image experiments show that it has high accuracy,which provides a new solution for visual tasks in specific situations in orchard environments.(2)The basic principle of YOLOv5 algorithm is introduced,and an improved YOLOv5 lightweight network model is proposed.The specific improvement strategy is as follows: the K-means++ algorithm is used to recluster the anchor boxes to obtain the most suitable anchor frame size for the citrus dataset,and the backbone network is replaced with the Mobile Net V3 lightweight module.At the same time,in order to effectively alleviate the accuracy degradation problem caused by model simplification,the CBAM attention mechanism module and Si LU activation function are introduced,and the loss function for single-class object detection is improved,and finally the improved MC＿YOLOv5 model is obtained.Experiments show that the MC＿YOLOv5 model designed in this paper reduces the m AP by 2.6%,reduces the model size by 80%,and increases the detection speed by 24% and 59% on the GPU and CPU platforms,respectively,which can better meet the recognition and detection of citrus targets in orchards.(3)Through binocular stereoscopic calibration and stereo correction,distortion is eliminated and the corresponding feature points of the left and right images are aligned with the corresponding feature points.Comparing the semi-global block matching(SGBM)algorithm with the block matching(BM)algorithm,it is found that the former performs better in terms of comprehensive performance.The positioning method in this paper fully meets the accuracy requirements of the picking robot for the three-dimensional spatial positioning of citrus targets,and the experimental data show that the average positioning error sizes in the three directions of X,Y,and Z axes are 4.27 mm,3.11 mm and 1.1mm,respectively,and the positioning error rates are within a reasonable range.