The Optimization Of ORB-SLAM System Based On Loop Closure Detection Technology

SLAM is the technology of synchronous orientation and map construction.It is a key technology for robots to solve positioning,map construction and navigation problems in an unknown environment.Among multiple SLAM systems,ORB-SLAM uses ORB features for tracking,mapping and location recognition tasks.It has received extensive attention on account of good instantaneity and stability.The complete ORB-SLAM system includes four links:tracking,local mapping,closed-loop detection and global optimization.The closed-loop detection can correct the drift error by detecting closed loop pairs in video frames,which is the key link in the SLAM system.The current closed-loop detection adopts the traditional ORB feature,and generates the feature vector of the image by constructing a thorough visual word bag.This method is defective in making spatial description of the image features,which is easily affected by environmental changes.In addition,the traditional word bag is too big to migrate to small mobile devices,which limits the application scenario of SLAM to some extent.Based on the above background,we combine the closed loop detection in SLAM with deep learning theory to optimize the SLAM system.The major research includes the following points:Firstly,we propose an unsupervised feature extraction method by using stacked hybrid automatic encoder after doing research on general image retrieval technology.On that basis,we establish a depth feature descriptor of multi-layer to optimize the extraction of feature.Other than this,a specific target detection algorithm is introduced in preprocess stage to identify the potential interference such as human in the dynamic images,which is helpful to improve the robustness of the model.Apart from feature extraction,the second research emphasis is how to improve the efficiency of similar image retrieval.Based on the theory of approximate nearest neighbor search,we choose hash algorithm as the feature coding method,and introduce the quantization error term in the learning of hash function.Hash coding can improve the efficiency of closed loop detection and meet the real-time requirements of the system.Finally,a secondary retrieval strategy is designed in order to make further improvement on the retrieval precision.After representing the image by hash code,we use clustering algorithm to find the neighbor images of the target image,and then calculate the similarity by the original feature code.This method has a greater advantage in efficiency than ordinary sequential retrieval.In this thesis,the performance test of each module of the system is carried out,and the results of experiments are analyzed.The results demonstrate the effectiveness of the method described in this paper.