Research On Farmland Obstacle Detection And Recognition System

With the development of driverless technology,automatic navigation technology has become the key technology of agricultural machinery automation and intelligent development.Autonomous Navigation of agricultural machinery operation can effectively reduce the labor intensity of workers,improve the efficiency and quality of agriculture,make agricultural machinery develop to a more accurate,automatic and intelligent direction,what can develop China’s modem agriculture rapidly.Among them,the detection and recognition of obstacles plays an important role in the automatic navigation and obstacle avoidance system of unmanned agricultural machinery.Therefore,it is of great significance to study the obstacle detection and recognition system of driverless agricultural machinery.In this study,the semi enclosed farmland is taken as the operation scene.By analyzing the characteristics of obstacles and comparing the applicability,advantages and disadvantages of existing environmental sensing sensors,the single line lidar and monocular camera are selected to design the farmland obstacle detection and identification system to detect and identify the location,distance,width,quantity and category information of obstacles.For the lidar module,this paper uses the median filter and nearest neighbor clustering algorithm to complete the screening and processing of LIDAR point cloud data,then designs and develops a GUI which can observe the location,distance and quantity information of obstacles visually.For the machine vision module,collect 5423 photos of farming scenes,and make labels and datasets.Based on AIstudio Deep Learning Platform,the datasets are trained with yolov5m and yolov51 preprocessing models respectively by using cloud GPU.Then compared and analyzed the recognition effects and speed of the two models,Yolov5m is selected as the system pretreatment model;Finally,the indexes of the training model are evaluated,the box loss is 0.0217,the objectness loss is 0.0182,the classification loss is 0.00025,the recall is stable at about 0.827,the precision is stable at about 0.913,and the Map0.5 reaches 0.865,mAP@0.5:0.95 reaches 0.606.According to the evaluation index,the qualitative analysis model has good detection and identification ability.Finally,an experiment is designed to verify the accuracy and effectiveness of the obstacle detection and recognition system.Experiments show that the filtering and clustering effects proposed in this paper meet the set requirements,and the single line lidar can effectively detect the location,size and quantity of obstacles;The monocular camera has a good effect on target detection at different distances and different orientations,and it can distinguish the types of obstacles accurately,and the recognition confidence is above 0.85.In conclusion,both lidar and machine vision can accurately meet the requirements of obstacle detection and recognition.