Research On Object Detection Method Based On Mobile Robot

In recent years,mobile robots have become increasingly powerful,which has benefited transportation,medical and industrial fields.Due to the increasing complexity of real-life environments,mobile robots need to cope with interference from factors such as lighting,occlusion,and size,so that human performance requirements for mobile robots are also increasing.In order to enable mobile robots to better understand their environment and provide a basis for judgment of subsequent movements,the research on target detection methods based on mobile robots has become a hot topic in current mobile robots.Improve the detection efficiency and robustness of mobile robots.To this end,the design and research of the mobile robot target detection system in the street view environment.Mainly through the research and combing of the technological development of mobile robots,deep learning and target detection,to understand relevant knowledge and to analyze the research significance and background of the current mobile robot-based target detection methods.Introduce and analyze the current single-stage target detection representative algorithm based on deep learning YOLO and the two-stage target detection representative algorithm Faster R-CNN.By considering the advantages and disadvantages of the two algorithms,Faster R-CNN is selected as the basic of mobile robot target detection system algorithm.Secondly,through the implementation of the two-stage target detection algorithm Faster R-CNN,to make corresponding improvements to its current detection performance deficiencies,first use deep linear convolutional neural network to fully extract target features;and then target difficult-to-detect vehicles For the target,introduce the idea of difficult case mining in the regional recommendation network to make the training more adequate,and use the clustering algorithm to determine the length-width ratio of the recommendation frame;in addition,for the problem of detecting small targets at a distance,the bilinear interpolation RoI regression is introduced in the pooling layer of the target area One algorithm;Finally,compare the parameter optimization algorithms commonly used in deep learning,and select the most suitable algorithm for this article.The five-point improvement strategies were experimentally verified to prove their effectiveness and feasibility.Finally,this paper designs and implements a target detection system based on mobile robots.In the Xicheng campus of Beijing University of Architecture,images are collected by a monocular camera,passed to the main control board for recognition,and finally transmitted to the display.Aiming at the problems of occlusion and small target scale,experiments on real scenes were conducted to verify the feasibility of mobile robot target detection system.