Research On Visual SLAM In Complex Dynamic Scene

With the development of related technologies such as artificial intelligence and computer vision,robots have gradually become one of the iconic products of the technological society.Visual SLAM is a prerequisite for robots to complete autonomous movement and carry out high-level tasks,and has great research significance in scientific research and industrial production.Indirect visual SLAM has gradually become the mainstream due to its stable operation,and its application in simple static scenes has become mature,but it is still in the exploratory stage under complex dynamic scenes such as dynamics,low textures,and obvious changes in light viewing angles.On the one hand,dynamic targets will cause mismatch or can not match of feature points,which reducing the accuracy of localization and the effect of mapping.On the other hand,most of the traditional image feature points used by visual SLAM are based on artificially designed key points and descriptors,and the matching performance is not enough to meet the needs in scenes such as low texture and obvious changes in light viewing angle.This thesis first studies the visual SLAM in dynamic scenarios.When there are too many dynamic targets,the SLAM localization accuracy is low,and it may even fail.This thesis first proposes a dynamic target detection algorithm of semantic segmentation and geometric tight coupling,using Mask R-CNN network to generate semantic mask,and using it as prior information to improve the accuracy of geometric motion segmentation and generate similar geometric masks.On this basis,this thesis integrates the dynamic target detection algorithm into the ORB-SLAM2 system,performs detection on the input image,initializes the weights of feature points based on the detection results,and jointly optimizes the weights and estimates the pose.In order to evaluate the performance of the dynamic SLAM system,this thesis conducts experiments on the TUM RGB-D dataset and real scenes.The results of qualitative and quantitative experiments show that the localization accuracy of this dynamic SLAM system in dynamic scenarios is significantly improved compared to ORB-SLAM2,and it is more robust than the current advanced dynamic SLAM algorithm.Aiming at the shortcomings of traditional image feature points in complex scenes,the advantages of automatic extraction of image features by deep learning are used to extract feature points and estimate the relative pose of the camera in a learning manner.Based on the self-supervised learning feature extraction network Super Point,the Soblel edge detection filter is used to obtain a gradient-guided feature detector,and then the feature detector and the Deep F pose estimation network are connected together and trained end-to-end.In order to achieve end-to-end training,this thesis implements a local soft argmax(softargmax)operation on the output of the feature detector to ensure that the gradient is passed back to the feature point.Finally,perform feature extraction and relative pose estimation experiments on multiple complex scene data sets.Feature extraction experiments under the Hpatches data set show that the learned feature points are more evenly distributed than traditional feature points,and can obtain higher matching scores.In the relative pose estimation experiment under the KITTI dataset,the estimated pose error of the algorithm in this thesis is the smallest;under the Apollo Scape dataset,although the pose error in this thesis is slightly larger than SIFT,it is relatively small compared to Super Point,which verifies that the end-to-end training method is correct Improvement of accuracy and generalization ability.