Research On Loop Closure Detection Algorithm For Visual SLAM Based On Deep Learning

Loop closure detection is a key technology for visual SLAM system to obtain global consistent map and trajectory by searching the loop closure in the trajectory to correct the trajectory error.However,it is difficult to guarantee the accuracy of loop closure detection based on traditional feature points in complex scenes with strong changes in illumination and seasonal changes.Existing loop closure detection algorithms based on deep learning can deal with complex scenes by relying on the strong representation ability of convolutional features to images.However,due to the complex model structure of convolutional networks,the algorithm has a large amount of computation and poor real-time performance.In order to maintain the performance of loop closure detection algorithm based on deep learning and improve real-time performance,the following work is completed in this paper:(1)Aiming at the problem of poor real-time performance of scene description based on deep learning model,this paper proposes a scene description algorithm based on lightweight VGG16.This algorithm combines VGG16 with NetVLAD and uses the output VLAD features to achieve scene description.In order to improve the real-time performance of the algorithm,ThiNet was used to cut the VGG16 redundant filter and lightweight the network,so as to improve the speed of feature extraction.At the same time,PCA was used to reduce the output high-dimensional VLAD features to improve the speed of similarity calculation,thus greatly improving the real-time performance of the loop closure detection algorithm in scene description.(2)Aiming at the problem of high time complexity of feature matching and easy misjudgment of loop closure,this paper proposes a feature matching algorithm based on hierarchical navigable small world graph.In this algorithm,the hierarchical navigable small world map is constructed from the extracted VLAD features to speed up the establishment of scene description library and scene feature matching.At the same time,sequence analysis and sequence judgment processes are introduced to reduce the loop closure misjudgment by skipping the scenes of judging adjacent time frames and scenes that are not similar to the surrounding scenes,thus reducing the time complexity of the loop closure detection algorithm in feature matching and the loop closure misjudgment.(3)In order to further verify the overall performance of the proposed algorithm in SLAM system,the loop closure detection module of ORB-SLAM2 was used to replace the algorithm proposed in the paper,and a comparative test was carried out on the open data set TUM RGBD.The experimental results show that the improved trajectory effect is more consistent with the real trajectory.The absolute trajectory error and relative trajectory error are smaller,and the performance is better than the original ORB-SLAM2.