Research On Observation Mechanism Of Obstacles In Specific Operating Environment Based On Binocular Stereo Vision

In intelligent agriculture,it is an essential function that obstacle perception in the field operation environment for the agricultural robot when working in the field operation environment.The detection of obstacle information in the work space is that the prerequisite for many applications.Such as agricultural robot autonomous navigation,picking operations,field weeding,and environmental monitoring.There are a variety of obstacle detection methods in the field.Ultrasonic wave,millimeter wave radar and lidar are difficult to accurately identify the number and type of obstacles.Monocular vision has the problem of object image occlusion and it is difficult to obtain target depth information.The process of multi-eye visual image matching is complicated.It takes a long time and slow response;multi-sensor fusion has the problems of complex algorithm and large amount of calculation.In this thesis,a single sensor(binocular camera)is used as a tool to study the obstacle perception mechanism in a specific operating environment in the field,and provide theoretical methods and experimental data for the application of obstacle perception in specific field operating environments.The research contents and related work arrangements of this thesis are as follows.(1)Analysis of imaging principle of perspective camera and calibration of binocular camera.The imaging principle of the perspective camera is analyzed,and the transformation formulas between the world coordinate system,camera coordinate system,image coordinate system and pixel coordinate system are calculated.With the help of MATLAB software,the internal parameters of the left and right cameras are obtained.Study the principle of binocular stereo imaging,solve the external parameters of binocular camera,and correct image distortion.The stereo calibration and correction research of the binocular camera has laid a hardware foundation for the subsequent binocular parallax method solution and the spatial distance of obstacles in the field.(2)Characteristic analysis of typical field operating environment obstacles.Compare the characteristics of mean filter,median filter and Gaussian filter in the image preprocessing of obstacles,choose median filter as the image preprocessing method;analyze by color histogram,color moment,color aggregation vector and color correlation graph image color features;use the structural method to analyze the texture features of image to obstacles,use Canny edge detection operators to obtain target edges of image,and use the Hough transform formula to express the shape of obstacles.The above-mentioned typical obstacle feature analysis and obstacle classification method research provide a theoretical basis for using deep learning recognition algorithms to identify obstacles.(3)Classification of obstacle image targets based on deep learning target recognition.Analyze the frame structure of the deep learning target recognition algorithm,verify the feasibility of the YOLOv3 algorithm for image recognition of field obstacles,cattle and other animals,and use the improved non-iterative K-means algorithm to achieve obstacle classification.Quickly realize the classification of non-obstacle targets and obstacle targets,and solve the problem of obstacle recognition classification based on target recognition.(4)Field obstacle ranging and three-dimensional space area construction.Using the binocular stereo vision parallax method to measure the depth of obstacle information theory,the feasibility of introducing binocular parallax matching ranging in the obstacle target recognition algorithm is analyzed,and discussed the question that pathological search and time-consuming of the local stereo matching and global stereo matching algorithms longer.Based on the epipolar constraint theory,a homology recognition frame matching algorithm is designed.Obstacle distance data is obtained from the center point of the homology recognition frame,and the spatial three-dimensional area of the obstacle is constructed according to the size of the recognition frame.The innovation of the work in this thesis is that reflected in the use of a single sensor(binocular camera)to achieve recognition and target classification,and the use of non-iterative K-means algorithm to determine the type of obstacle.And design of the algorithm that matching ranging based on homology recognition frame.After that use of field obstacle recognition frame size,reproduce the area of the three-dimensional space.