Design Of Obstacle Detection System For Orchard Robot Based On Object Detection Neural Network

The application of orchard robots in agriculture has made new breakthroughs in recent years,and its unmanned and automatic operation mode has set off a boom in the world’s agricultural field.In the process of operation,orchard robots will inevitably encounter some obstacles.Being able to accurately and quickly identify these objects and respond to them is the core issue of exploring the development and application of orchard robots.In recent years,deep learning research has developed rapidly,target detection methods based on convolution neural network are changing rapidly.In this dissertation,the position location and classification recognition of obstacles in orchard are realized through the deep learning method of visual perception,which provides strong support for the development and application of obstacle detection.The main research contents are as follows:(1)The test platform was built and the perceptual control program was designed.ZED binocular stereo camera was used to shoot video stream data of orchard scene,obtain real-time information of orchard,and process video stream data.Aiming at the acquired motion fuzzy image,the removal of fuzzy operation is carried out,and the obstacles in the image after the removal of fuzzy location and classification recognition,so as to obtain the accurate coordinates and category information of obstacles.(2)An improved YOLOv5 s target detection model based on deep learning convolution neural network is proposed,and the ZED binocular stereo camera is used to identify and locate obstacles in the orchard.Improve the CSPDarknet used in the backbone network of YOLOv5 s original model,and the common convolution is replaced by the deep separable convolution.After the improvement,the network computation and parameters are actually reduced by about 2/3.The CIo U＿Loss is used as the loss function of bounding box regression,replacing the original GIo U＿Loss improves the speed and accuracy of prediction box regression.The performance of the improved YOLOv5 s target detection algorithm is verified by the self-made orchard obstacle dataset,and the Precision is 0.65% higher than the original YOLOv5 s model and 0.83% higher than YOLOv3 model respectively.The Recall is0.81% higher than the original YOLOv5 s model,0.29% higher than YOLOv3 model,The detection speed is 333.74% faster than Faster RCNN,13.21% faster than the original YOLOv5 s model,and the parameter quantity is 97.26% less than Faster RCNN,29.17% less than the original YOLOv5 s model.Under the condition that the near and medium distance detection effect is similar,the improved YOLOv5 s model has better detection effect for long-distance dense small targets.(3)In order to solve the problem that the orchard robot can not accurately and quickly detect obstacles due to the input of blurred images during operation,the Deblur GAN-v2 image deblurring algorithm based on the GAN(Generative Adversarial Network)is selected and fused with the YOLOv5 s target detection algorithm,and a D2-YOLO one-stage deblurring recognition network based on fusion improvement is proposed.Through test verification,the time for D2-YOLO deblurring is about 1/10 of SRN-Deblur and about 1/3 of SSRN-Deblur Net.When do the deblurring obstacle recognition and detection,the accuracy and recall rate of D2-YOLO detection are 91.33% and 89.12% respectively,which are 1.36% and 2.7%higher than the Deblur GAN-v2+YOLOv5s of step-by-step training,and 9.54% and9.99% higher than the YOLOv5 s original model,which can effectively meet the accuracy and real-time requirements of robot obstacle deblurring detection in orchard scenes.