Research On Obstacle Recognition And Obstacle Avoidance Technology Of Tracked Robot Based On Binocular Vision

The rapid development of agricultural robots has greatly improved the efficiency of agricultural production and reduced the labor intensity of farmers.However,in order to cope with possible complex situations,agricultural robots must have the ability of obstacle detection and recognition and the ability of autonomous obstacle avoidance when operating independently in complex farmland environments.This is a problem that must be faced in the development of agricultural robots.Therefore,it is of great significance to study the obstacle detection and obstacle avoidance strategies for the development of agricultural robots and intelligent agriculture.Aiming at the possible situations in complex farmland environment,this paper proposes an improved obstacle detection algorithm based on yolov5 and studies the obstacle avoidance strategy based on binocular vision.The main research work of this paper is as follows:(1)The principle of imaging and ranging of binocular camera is analyzed.The binocular camera is calibrated by matlab toolbox.According to these parameters obtained by calibration,the binocular camera is corrected and the binocular camera is matched by SGM stereo matching algorithm.(2)The images of three types of farmland obstacles,including human,agricultural machinery and sheep,were collected.Through left and right flipping,random scaling,random addition of shelters,brightening and darkening processing,the scene of illumination change,shooting angle,near and far change and occlusion in the field was simulated.The training set of the data set was extended to 15000 sheets and marked by Label Img labeling software to make a standard PASCAL VOC 2007 format data set.(3)The obstacle detection algorithm based on yolov5 is studied,and the input end,Backbone layer,Neck layer and Head layer of yolov5 are analyzed.yolov5 s is selected as the benchmark model,and the model is trained by using the produced data set.The network weight obtained after model training is 14.4mb,m AP @ 0.5 reaches 87.7 %,recall rate reaches 83.1 %,and accuracy rate reaches 87.0 %.The average accuracy of sheep reached90.2 %,the average accuracy of human reached 86.1 %,and the average accuracy of agricultural machinery reached 87.2 %.(4)Aiming at some problems of detecting target obstacles in complex farmland environment,the yolov5 algorithm is improved.The new feature fusion network and the improvement of multi-probe design are introduced,the k-means clustering algorithm is optimized,the CBAM attention mechanism is introduced,and the ghost convolution is used to improve the model.It was verified by ablation test and comparative test.The ablation test shows that the average detection accuracy of the improved yolov5 algorithm is 2.3 % higher than that of yolov5 s,the recall rate is increased by 3.1 %,the accuracy rate is increased by1.9 %,and the number of parameters is reduced by about 7 %.The SSD,Faster RCNN,yolov3,yolov4,yolov5 s and yolov5 improved algorithms are compared.The experimental results show that the improved yolov5 algorithm has the highest average accuracy and the fastest detection speed.It further proves the advantages of yolov5 improved algorithm and achieves the expected goal.(5)Study the artificial potential field method,point out the existing defects and make improvements,and use matlab software to carry out simulation test verification.The test results show that the improved artificial potential field method can make the robot avoid obstacles smoothly and reach the target point.(6)The field test was carried out,and the feasibility of the system was verified by setting up obstacles.