Guava Detection And Location In Natural Environments

The picking of target fruits by agricultural intelligent robots in complex environments has always been one of the most concerned problems in the industry.In order to promote the industrialization of fruit picking robot and improve the comprehensive performance and application value of fruit picking robot,this paper combines classical machine learning method,convolution neural network theory and binocular stereo vision technology,takes guava in natural environment as the research object,studies image acquisition,image preprocessing,accurate segmentation,recognition and positioning of guava,and develops guava.Intelligent recognition and positioning system of pomegranate provides an effective and reliable technical support for picking guava or similar fruits in complex environment.The main research work of this paper is as follows:1)Collection of guava image.In natural environment,the camera is calibrated,including single target calibration and stereo calibration,and the internal and external parameters of the camera and the structural parameters between the cameras are obtained.Then the image of guava is collected by binocular stereo vision system.The distance between the left and right cameras and guava is about 500 mm,totaling 500 images,and the captured image is converted into two-dimensional matrix and stored in notebook.In the computer.2)The accurate segmentation and recognition of guava was studied.In the natural environment,it is difficult to separate the target fruit from the background by the traditional color model segmentation method,which is affected by the illumination change.In this paper,texture features are selected to distinguish guava from background in order to segment guava accurately.Aiming at the low classification accuracy of single classifier,the weak classifier is constructed into strong classifier based on Ada Boost algorithm,and several strong classifiers are combined into cascade classifiers to improve the recognition accuracy of guava.Ada Boost cascade classifier based on MB-LBP features was used to detect and recognize guava in 64 test images.The experimental results show that the recognition accuracy of guava is 85.5%,the recall rate is 66.3%,and the average processing time of the image is 0.77 s.At the same time,the method based on full convolution neural network is used to segment and recognize the guava image.The features of multiple layers of the image are extracted automatically through multiple convolution layers and pooling layers,and the pixels in the image are classified one by one.The guava can be accurately segmented from the image to realize the classification at the image pixel level.The training full convolution neural network is used to segment the guava in the test image.The performance of the full convolution neural network is evaluated and analyzed from the indexes of pixel accuracy,average pixel accuracy and average intersection-to-merge ratio.The average pixel accuracy is 99.1%,average pixel accuracy is 99%,average intersection-to-merge ratio is 85.3%.3)The acquisition of guava centroid and the calculation of its three-dimensional coordinates were studied.Firstly,the basic camera imaging model,common coordinate system and their relationship are introduced.Then,using the calibrated camera parameters and the structural parameters between the cameras,the segmented left and right image pairs are calibrated to align the left and right image pairs.Then,the minimum outer rectangle of the left and right image pairs for guava is found and the serial number is labeled.The center of the rectangle is taken as guava.The center of mass.According to the sequence number,the center of mass of guava is matched and the parallax is calculated.The three-dimensional coordinates of the center of mass of guava can be obtained by combining the principle of triangulation.Finally,the measurement error test of the binocular vision system is carried out.The experimental results show that the measurement error of the binocular vision system is between-0.06 mm and 0.08 mm,and the maximum absolute error is not more than 0.12 mm.That is to say,the measurement accuracy of the binocular vision system selected in this paper is high.According to the above theory,the three-dimensional coordinates of guava centroid can be obtained.That is to say,the positioning method of guava presented in this paper is feasible.4)An intelligent segmentation recognition and positioning system for guava was developed.In this paper,a guava intelligent segmentation recognition and positioning system is designed and a human-computer interaction interface is made.The method of guava segmentation recognition and positioning used in this paper is programmed to verify the feasibility of the method of recognition and positioning used in this paper,which can meet the needs of fruit picking robots in natural environment.