Research On Object Detection In Lawn Environment Of Embedded Mowing Robot Based On Deep Learning

With the vigorous development of intelligent agriculture with deep learning as the core technology and the commercialization of lawn mowing robots,the requirements for the visual inspection system of lawn mower robots are getting more and more higher.The embedded computing device on the lawn mowing robot needs to visually perceive the lawn trees and pedestrians and make decision for control based on the information collected by the camera and other visual sensors.Therefore,the detection accuracy and lightweight degree of the lawn target in the mowing scene have very high requirements.Aiming at the current lack of public lawn environment object datasets and the low detection accuracy and low degree of lightweight of the original Tiny YOLO series of object detection algorithms,this thesis establishes a lawn environment object dataset and proposes an embedded object detection scheme based on deep learning.The main work is as follows:A dataset of objects for lawn environment has been produced.Smart phone camera is used to take pictures of a large number of lawn trees in school lawns,park lawns,and golf lawn environments,filtering out samples containing major lawn tree types.Use Label Img to mark them to form a label file in YOLO format to form a lawn tree dataset,including trunk and spherical tree.The person dataset in the PASCAL VOC2007 dataset is extracted,combined with the lawn tree dataset to form the lawn environment object dataset,and divided into training set,validation set and testing set according to the ratio of 16:4:5.The kmeans++algorithm is used to cluster six types of a priori boxes.Aiming at the problem that the Tiny YOLOv3 algorithm does not fully utilize the shallow information of the network and the detection accuracy is low,the Optimized Tiny YOLOv3 algorithm is proposed.The method of channel scaling and multi-resolution fusion improves the utilization of shallow information while reducing the amount of calculation.Experiments show that the amount of calculation is reduced by 3.67%,and the m AP(mean average precision)value is increased by 8.49%.Aiming at the problem of too many convolution parameters of the Tiny YOLOv4 algorithm,an optimized Tiny YOLOv4 algorithm is proposed.Through the channel multi-scale transformation,the enhancement module and the lightweight module are designed to replace the zoom module;the multi-resolution fusion module is designed through the principle of multi-resolution fusion,which replaces the last convolutional layer of the the backbone.Experiments show that the m AP value is increased by 1.35%,and the number of convolution parameters is reduced by 27.94%.A high-precision and lightweight Embedded YOLO lawn environment object detection algorithm is proposed.Based on the channel incremental depth convolution and residual suppression theory,the Embedded-A module is proposed,which expands the depth of the feature map twice and forms residuals to improve the lightweight degree.Based on the theory of residual fusion,the Embedded-B module is proposed to improve the accuracy of feature map downsampling through deep convolution and pooling fusion.The Embedded YOLO object detection network is formed by the stacking of Embedded modules and the fusion of feature maps of different resolutions.Experiments show that the Embedded YOLO algorithm has a m AP value of 77.20%on the testing set of the lawn environment object dataset,the number of convolution parameters is only 1.78(?)10~6,the calculation amount is only 3.85BFLOPs,and the weight file size is only 7.11MB.The lawn environment object detection algorithms proposed in this thesis all improve the detection accuracy while greatly improving the degree of lightweight.It provides a theoretical basis for the application of computer vision technology based on deep learning to embedded lawn mowing robots.