Research On SLAM Loop Closure Detection Based On Siamese Convolutional Neural Network

SLAM(Simultaneous Localization and Mapping)technology is a comprehensive technology that integrates multiple functions such as perception,control and mapping,which allows robots to navigate and locate autonomously in unknown environments,and has been widely used in robotics,autonomous driving and other fields.Two types of algorithms for SLAM can be roughly distinguished: filter-based and image-based.Since traditional filter-based methods,such as Kalman filter and particle filter,accumulate errors when building maps,they are not suitable for large-scale map creation.Therefore,recent research mainly focuses on image-based methods.Loop closure detection is a key challenge in imagebased visual SLAM tasks.It refers to determining whether a robot has visited a previously reached location,and by correcting for accumulated errors to generate a globally consistent map.The traditional method of loopback detection is very sensitive to changes in the external environment.In response to this problem,the method of deep learning has gradually been used in the loopback detection problem.The SLAM loop closure detection method based on the siamese convolutional neural network proposed in this thesis,the main research contents include:(1)Aiming at the problem that RGB images in complex scenes are greatly affected by external conditions such as illumination,viewing angle changes,dynamic objects,and weather,this thesis introduces a deep learning method to extract features from scene images.On this basis,in order to perform loop closure detection more efficiently,this thesis proposes a method based on the siamese convolutional neural network,which can extract the features of the same distribution domain of the two pictures,so as to better compare the similarity of the two scenes.(2)For the problem of loop closure detection in complex scenes,this thesis uses VGG16 and Alex Net as the sub-feature extraction network of the siamese neural network.Using VGG16 combined with the siamese neural network,and trying to use the single-task learning method to supervise the network training,and finally use the Sigmoid function to compare the scene similarity.Experiments prove that the loop closure detection method based on the siamese convolutional neural network is significantly better than the general deep learning method and the traditional method of manually constructing features.In order to give full play to the advantages of the siamese neural network,this thesis improves the method based on the automatic learning of feature descriptors by Hou et al.using Alex Net,using Alex Net as a sub-feature extraction network of the siamese neural network,and using multi-task learning.By employing two distinct loss functions to oversee the network training,and then utilizing the stochastic gradient descent technique to update the network parameters,the accuracy of the loop closure detection can be enhanced.This thesis refers to VGG16-based siamese neural networks and Alex Netbased siamese neural networks as siamese convolutional neural networks.(3)This thesis employs a plethora of training and test sets to enhance the precision of similarity comparison in loop closure detection.For the two methods of VGG16 and Alex Net,the Sigmoid function and L2 norm distance are used for scene similarity.Judgment,and finally use the P-R curve to display the accuracy of the network.The implementation results show that the SLAM loop closure detection method based on the siamese convolutional neural network is significantly better than the traditional method of manually constructing features,and is also better than the general deep learning method,which lays the foundation for mobile robots to perform tasks in more complex scenes in the future.