Object Detection Based On Lightweight Neural Network In Mobile Robot

Object detection is one of the most important ways for mobile robots to perceive the environ-ment.Mobile robots can use their own cameras to acquire images of the surrounding environment.Object detection uses images to help robots detect and locate potential targets.This thesis focus-es on applying object detection algorithms based on deep learning to mobile robots.In order to solve the problem that the computing resources of mobile robots are limited,lightweight backbone networks are used in all object detection networks in this thesis.This thesis implements three types of object detection algorithms to solve multiple problems in mobile robots.Three types of object detection algorithms are two-dimensional object detection,rotating object detection and three-dimensional object detection.Mobile robots include quadrotor and mobile shooting robot.Four issues are:two-dimensional object detection in quadrotor obliquely down-view images,rotat-ing object detection in quadrotor down-view images,autonomous target strike in mobile shooting robot and monocular three-dimensional target localization in mobile shooting robot.Above all,this thesis includes the following four parts:1.Two-dimensional object detection in quadrotor obliquely down-view images.Faster-RCNN is improved and a lightweight backbone network is designed to speed up the running speed of the whole algorithm.While ensuring detection accuracy,the lightweight backbone network solves the problem that the original Faster-RCNN algorithm can not run at high speed in real time.According to the fact that the sizes of targets in the oblique down-view image are relative to targets' ordinates in the image,an incremental anchor is proposed in this thesis.The size of the incremental anchor is linearly related to its ordinate.It is proved that the proposed incremental anchor can effectively improve the accuracy of detection algorithm.The proposed algorithm is integrated to an environmental percetion module which helps quadrotors autonomously search for ground mobile robots.This work is validated in the International Aerial Robotics Competition.2.Rotating object detection in quadrotor down-view images.A two-dimensional rotating object detection method based on deep learning is proposed.The object detection network mainly includes three modules:a lightweight backbone network for extracting image features and generating a feature pyramid;a rotation subnetwork that performs five-parameter regression on the feature pyramid,acquires rotating bounding boxes and categories of targets;an head regression module which classifies the heads or tails of targets separately to obtain targets'heading direction.Experiment shows that the proposed method can effectively detect the rotating targets while estimating their the heading direction.3.Autonomous target strike in mobile shooting robots.A solution based on deep learning for target strike is proposed in this thesis.The proposed solution mainly includes two parts:Firstly,a lightweight backbone network is designed so that the improved object detection algorithm can run 26 frames of images per second on a mobile computing device TX2 while maintaining high detection accuracy.Secondly,an autonomous target strike method is pro-posed for mobile shoot robots.Object detection results are used to control the two-degree-of-freedom turret to shoot at moving targets.A state machine is designed to control the bullet launching machine to solve the problem that the camera's perception range is too small.The overall proposed method is validate in experiments and the procedure that the mobile robot successfully shoots at moving targets is showed.This work is validated in the 2019 ICRA Robomaster Artificial Intelligence Challenge4.Monocular three-dimensional target localization in mobile shooting robots.M3D-RPN is used for three-dimensional target localization in mobile shooting robot.A longitudinally separated convolution operation which can effectively improve target localization accuracy is proposed.Experiments show the proposed method can effectively localize mobile shooting robots.