| Agriculture as the foundation of China’s economic development,agricultural mechanization occupies a key position in its modernization,so in this form,people focus on agricultural machinery.This paper focuses on the early research work of fruit picking machinery,and is committed to solving the problem that fruit recognition accuracy is not high due to insufficient light.In this paper,we first to set up the experimental platform,the mobile platform chassis to walk in the garden,safely and accurately to orchard environment built figure,for fruit trees and locations for perception and avoid the obstacles,to the fruit on the trees,at the same time especially in all kinds of light dark conditions needs to keep a certain accuracy.In this paper,lidar and RGBD depth camera are used to scan the environment and complete the autonomous navigation of the mobile platform chassis,such as map building and obstacle avoidance.Fruit recognition is accomplished by deep learning convolutional neural network.The details are as follows:(1)Briefly explain the research significance and background of this topic,investigate and analyze the status quo of autonomous navigation and fruit target recognition at home and abroad,compare the mainstream target detection algorithms,explain the reasons for choosing YOLOv4 for research,and briefly explain the current development trend.(2)When the orchard,based on the autonomous navigation system of the construction of the mobile platform chassis,including hardware system,laser sensor,vision sensors and ROS operating system,overall architecture of autonomous navigation system,complete chassis movement control and correction,the braking performance of experiment under different scenarios.(3)Focus on the synchronous positioning and mapping.The main ALGORITHMS of SLAM such as Gmapping,Hector-SLAM,Karto-SLAM,Cartographer and RTABMap are used to independently build maps in the indoor environment and conduct comparative experiments.According to the integrity of the built image and other aspects of the analysis of the experimental results,r TAB-Map was selected to carry out the following outdoor building map,the experiment,the path planning adopted global and local planning combination method,after comparative study,the former used A*algorithm,the latter used TEB algorithm,indoor and outdoor experiments show that,Mobile platform chassis can complete basic autonomous obstacle avoidance navigation requirements,laying a good foundation for the following fruit recognition.(4)The convolutional neural network and the target detection,on the basis of YOLOv4,aiming at environmental particularity,in network structure on the basis of YOLOv4 CBAM(convolution attention mechanism module)and gradient adaptive illumination regulation network model,improve the model of bad light the accuracy of target recognition and feature extraction ability.The new improved network model is named SLA-YOLOV4.Data sets were collected and comparative training experiments were conducted with YOLOv4 and SLA-YOLOV4 respectively.Under daylight conditions,the accuracy of SLA-YOLOV4 was improved by 1.06% compared with YOLOv4,and the overall recognition rate was improved by 2.62%.Compared with YOLOv4,SLA-YOLOV4 improved the accuracy by 3.79% and the overall recognition rate by 6.76% under night conditions.This indicates that SLA-YOLOV4 model has a certain improvement effect,and the effect is better at night.It was deployed on the mobile platform Jetson Nano board through Tensor RT,and the simulation experiments of mobile platform map construction,fixed-point navigation and fruit recognition were completed in an outdoor simulated orchard environment,which verified the feasibility of the proposed scheme. |
PDF Full Text Request