Research On Loop Closure Detection Of Visual SLAM Based On Deep Learning In Dynamic Environment

With the rapid development of artificial intelligence technology,researchers have put forward high requirements for autonomous positioning and mapping of mobile robots.Simultaneously Localization and Mapping(SLAM)has attracted significant attention in robotics community due to its excellent performance in localization and mapping.Loop closure detection plays an important role in visual SLAM,which aims to reduce the accumulated position errors caused by visual odometer and construct a consistent map by judging whether the mobile robot has reached previously visited places.However,loop closure detection is still a challenging problem because two images of the same place taken at different time may change dramatically due to the interference from viewpoints and dynamic objects such as people and vehicles.In addition,although loop closure detection methods based on hand-crafted features perform well in simple scenes,they are still difficult to obtain good results in complex dynamic environments because they are highly dependent on the manual features.In recent years,deep learning has been widely used in various computer vision tasks,such as image classification and image retrieval,and tends to replace hand-crafted features due to its outstanding feature representation ability.In view of the development trend of loop closure detection and the powerful feature representation capability of convolutional neural network,we propose two loop closure detection algorithms based on deep learning in this paper.Firstly,an online loop closure detection algorithm based on semantic and appearance features is proposed.Semantic segmentation network Deep Lab V3 splits out common dynamic objects of images and then convolutional neural network VGG16 is used to extract global feature of the remaining static background image.For the global feature of input image,hierarchical navigable small world is employed to retrieve the loop closure candidate of input image.Finally,local difference binary descriptors and RANSAC are employed to verify the loop closure matches.Compared with other typical or state-of-art loop closure detection algorithms,the proposed method can achieve higher recall rate and efficiency on several public datasets.Secondly,a lightweight loop closure detection algorithm based on second-order attention and Net VLAD is proposed.The crucial component of the algorithm is a network with Efficient Net-B0 as backbone for global feature extraction,which integrates second-order attention module and Net VLAD layer.The second-order attention module is applied to learn the correlation between features within the feature map,and the Net VLAD layer is employed to generate a compact and fixed-length global feature.Knowledge distillation strategy is adopted in training of the proposed network to speed up the training process.For the global feature of input image,hierarchical navigable small world is also applied to retrieve the corresponding loop closure candidate image.Geometrical consistency check based on local difference binary descriptors and RANSAC is designed to verify loop closure matches.Experiments on several public datasets show that the proposed loop closure detection algorithm can obtain higher recall rate under 100% precision compared to other typical and state-of-the-art methods.The global feature extraction network can effectively obtain the correlation between features within the feature map,and improve the ability of feature representation.