| Simultaneous Localization and Mapping(SLAM)is a key technology that enables robots to gain autonomous perception and navigation in unknown environments.SLAM can also provides important support for the future development of intelligent transportation,smart logistics and autonomous driving.Current vision SLAM technology based on the traditional manual key points methods have made a lot of achievements,but the feature extraction methods based on traditional manual features can have the problem of not being able to extract enough key points correctly or having low matching accuracy in scenes with complex changes in lighting,seasons and other changes.This problem can lead to visual SLAM systems that are unable to perform pose estimation and loopclosing detection in complex and variable environments.With the development of deep learning technology,current Convolutional Neural Networks(CNN)based algorithms can be a good solution to the problems of manual feature extraction methods in changing and complex scenarios.In visual SLAM,loopclosing detection techniques can reduce the effect of cumulative drift and improve the global consistency and accuracy of the map.Traditional loopclosing detection methods are less robust to environmental changes,while current convolutional neural network-based loopclosing detection techniques,although improved in robustness,are difficult to persist,reload the database,and also difficult to integrate with current mainstream visual SLAM systems.Therefore,the main work of this thesis are as follows:Dimensionality reduction of neural network feature descriptors using image global descriptors and Principal Components Analysis(PCA).In this thesis,the dimensionality reduction of feature descriptions not only improves the computational and storage efficiency,but also achieves high matching efficiency and average matching accuracy at the cost of losing some key points.Design a location recognition algorithm based on Bag of Words(BoW)model and convolutional neural network features,and the location recognition algorithm obtains better prediction accuracy than the traditional location recognition algorithms with manual features.The thesis also uses this algorithm to improve the loopclosing detection algorithm,which can be adapted to both complex scenarios with variations and can be applied to current mainstream visual SLAM frameworks.The algorithm outperforms traditional loopclosing detection algorithms based on artificial features in terms of loopclosing detection accuracy.In addition,the algorithm can directly use fisheye images for loopclosing detection,making full use of fisheye image information.Experiments show that more correct loopclosing locations can be detected using fisheye images than using fisheye corrected images.Finally,this thesis implements a binocular vision SLAM simulation system based on the improved feature extraction algorithm and loopclosing detection algorithm,which is based on SPTAM(Stereo Parallel Tracking and Mapping).The final experiments show that the results obtained from this binocular vision SLAM simulation system are closer to the ground truth than SPTAM and ORBSLAM2,and more robust to changes in the scene. |
