Research On Visual Loop Detection Algorithm In Complex Dynamic Environmen

Loop closure detection(LCD)is a key component of simultaneous localization and mapping(SLAM),which aims to detect whether a robot arrives a previously visited place.LCD can reduce accumulated errors by loop closure constraints of robot moving routes during SLAM.According to different sensors,LCD can be classified into camera-based LCD and Lidar-based LCD.Compared with Lidar data,visual data provides more information,but is heavily influenced by environment changes such as light,viewpoint,dynamic objects.Furthermore,the visual LCD algorithm relies on the comparison of the current image with the historical images.The time required for LCD will increase as the historical images accumulate,and the real-time requirement of loop detection cannot be met for large datasets or large scenes.Designing a robust LCD algorithm that meets the real-time requirement in complex dynamic environment is the research topic of this thesis.The main contents are listed as follows:(1)Loop closure detection by comparing image sequences in complex dynamic environment.The input data of the visual LCD algorithm is only the camera image,which contains both useful information(such as landmarks and fixed objects)and disturbance terms(such as changes of light,viewpoint,and moving objects).To improve the LCD algorithm,this thesis uses YOLO(You Only Look Once)algorithm to remove moving objects,extracts ORB(Oriented FAST and Rotated BRIEF)features from the rest of the image,and aggregates them into a low dimensional feature vector to represent the image.Furthermore,since changes of environment might cause dramatic changes in the image,we detect similar images by comparing sequences of images,rather than two single images.Experiments on publicly available datasets show that the proposed LCD could apparently improve the accuracy of detection and find out more potential loop closures.(2)Real-time loop closure detection with approximate nearest neighbors search algorithm.Applications of robot SLAM usually have a high real-time requirement,such as localization of self-driving vehicles.As the length of the route increases,the LCD algorithm has to match more images,which costs more searching time.To reduce the searching time,we use the Hierarchical Navigable Small World(HNSW)algorithm to search for matched images.By converting sequences of images into nodes in the HNSW graph,we could search the graph with complexity of O(7)log(7)n(8)(8).The LCD algorithm based on HNSW dramatically improves the real-time performance compared with the brute force algorithm.