Reserch On Visual Module Of Intelligent Mowing Robot Based On Deep Learning

Intelligent mowing robots can autonomously complete the mowing task and reduce a lot of unnecessary labor,so more and more people begin to carry out related research.At present,the mainstream intelligent mowing robot program at home and abroad is aimed at small and medium-sized lawns in a simple environment.The visual module is the most important part of the intelligent mowing robot.This dissertation attempts to research a smart mowing robot vision module for medium and large lawns with complex environment.The module has the function of identifying and locating obstacles.A multi-level obstacle extraction algorithm is proposed,for the problem that multiple obstacles can not be distinguished in the camera view,which effectively reduces the misjudgment problem when dealing with overlapping problems,and effectively solves the problem of high missed detection rate.A simple and efficient scene modeling method is designed,for the working scene of intelligent mowing robot.This method ignores the errors of image distortion and stereo matching,reduces the calculation amount and meets the precision requirements of mowing robots.The network structure based on the deep learning intelligent mowing robot obstacle recognition algorithm is designed.The multi-scale feature map is used for prediction.For the intelligent mowing robot working scene,the equidistant dense layout feature map is adopted to improve the detection accuracy.The loss function of the deep learning obstacle recognition algorithm uses softmax loss to represent the category confidence error.The multi-level positioning candidate frame is designed.The default box uses the information provided by the multi-level obstacle extraction algorithm to make the intelligent mowing robot have positioning obstacles.The mapping relationship between the obstacle bounding box and the world coordinate system is established.The PASCAL VOC format data set was created for the intelligent mowing robot working scene,and the data enhancement processing was used to increase the sample size.The common algorithm evaluation index was analyzed,and the accuracy rate was selected as the evaluation index of the intelligent mowing robot obstacle recognition algorithm.An effective training strategy was developed,gradient descent method was adopted as the training strategy in the initial stage of training,and then the Batch Gradient Descent method was adopted,which not only reduced the probability that the deep learning model did not converge,but also guaranteed the training speed.The hardware platform of intelligent mowing robot was built,and the parameter adjustment and on-site debugging of the mowing robot working scene were carried out.The accuracy of the obstacle recognition of the intelligent mowing robot vision module was tested with the accuracy rate and recall rate.The recall rate reached 99.03%,and the comprehensive accuracy rate reached 96.59%.The maximum ranging range,distance measurement accuracy,and obstacle size measurement accuracy were used as evaluation indexes.The obstacle positioning accuracy test was performed on the vision module of this paper.More than 10%,the farthest distance measurement reaches 11.32 m,and the obstacle size measurement width and height error are within 30 cm.