Research On Loop Closure Detection Of SLAM Based On Deep Learning

With the development of artificial intelligence,robots are widely used in various fields such as transportation,home furnishing,industry,and production.Simultaneous Localization and Mapping(SLAM)systems for mobile robots have become the focus of research by many scholars.Among them,the Loop Closure Detection algorithm,as an important part of the visual SLAM system,can provide effective constraints for the back-end pose optimization,thereby reducing the cumulative error caused by the motion estimation process,which is useful for realizing the synchronization of mobile robots in real time.Positioning and map construction play a decision-making role and are an indispensable part of the SLAM system.The traditional loop detection is realized by manually extracting features and constructing a visual bag tree model.The manual design of the method of extracting image features is too complicated and the uneven distribution of image features leads to poor final loop accuracy;the large-scale and complex calculation of the bag-of-words model ultimately leads to inefficient loop detection algorithms.Therefore,in response to the above problems,the research content of this article is as follows:(1)In this paper,combined with deep learning convolutional neural network related algorithms,the Goog LeNet and Res Net network frameworks are used to extract the deep features of real-time environmental vision images acquired by mobile robot camera sensors,and apply the ideas of "sparse connection" and "jump connection" into the SLAM loopback detection system.Contextual information is effectively added to the network,which can try to avoid the loss of characteristics when information is transmitted between layers.(2)The PCA+Whitening dimensionality reduction algorithm is used for the feature maps extracted by the Inception?V3 network to reduce the dimensionality of high-dimensional feature vectors,which greatly reduces the workload and improves the work efficiency of the SLAM system without loss of accuracy.For the feature map extracted by the Res Net V2?152 network,the TSNE dimensionality reduction algorithm is used to find the low-dimensional manifold in the high-dimensional space and reduce the dimensionality of the feature vector,which brings convenience to the calculation of similarity later.After selecting the Euclidean regularization method to calculate the similarity of the key frame image depth feature vector,the clustering algorithm of the Gaussian mixture model is applied to the similarity matrix,and the range of similarity matrix values is enlarged and eliminated the false detection point of the false positive loop.(3)Finally,using public environmental data sets for verification.Experiments show that the loop detection algorithm based on convolutional neural network,dimensionality reduction algorithm and Gaussian mixture clustering model can generate better image feature representation and effectively reduce the complexity of loop detection algorithm that manually extracts features and then performs feature matching,while improving the loop accuracy.(4)In this paper,an online medical registration system is implemented by using the algorithm of Res Net and TSNE.