Co-localization Technology Based On Convolutional Neural Network Object Detection

In the field of intelligent robot research,it is of great significance for intelligent robots to realize autonomous navigation and obstacle avoidance by using sensors to realize their own positioning and obtain accurate environmental characteristic information.At the same time,the higher the accuracy of positioning and map construction,the more accurate the intelligent robot can achieve functions such as positioning and navigation.In this thesis,the object detection algorithm based on convolutional neural network is combined with the SLAM system to achieve the improvement of SLAM accuracy.Aiming at the error of the input part of the system,the process of converting the points in the three-dimensional space to the imaging plane of the camera is analyzed,and the causes and correction methods of the distortion in the camera imaging are studied.Visual SLAM is mainly divided into visual odometer,back-end,and loop detection.The visual odometer can be divided into feature points and direct methods.Various characteristics and feature matching of ORB feature points are researched and analyzed,and experimental verification is carried out,and the data set is run through ORB-SLAM2.The advantages and disadvantages of traditional object detection algorithms and object detection algorithms based on convolutional neural networks are analyzed.The convolutional neural network is analyzed,and the indicators and performance of multiple object detection algorithms are evaluated.The object detection algorithm used in this topic is selected.And through the algorithm for experimental verification.Using the vanishing point obtained by the perspective principle and the twodimensional bounding box,a three-dimensional object detection algorithm is designed,and the corresponding loss function is designed to select a high-quality three-dimensional bounding box.Add the three-dimensional bounding box to the SLAM system,add new constraints on the basis of the original feature points and camera pose,and calculate the measurement errors of the three parts to obtain higher-quality poses and feature points,which are correlated through data And loop detection to achieve higher precision positioning and mapping.The construction of the system was completed on the Ubuntu system,and the environmental characteristic information mapping experiment was carried out using the KITTI data set.The absolute trajectory error of the system before and after the improvement was compared with the tool.The experimental results verified the effectiveness of the method in this thesis.