Research On Visual Loop Detection Under Complex Scene Changes Based On Deep Visual Perception

Vision SLAM refers to a robot equipped with a camera sensor,which autonomously estimates its own pose while building a map of the surrounding environment.In practical applications,the robot needs to continuously run in variable complex scenes for a long time.It requires the robot to have the ability to recognize scenes to adapt to complex scene changes,including illumination changes,season changes,occlusion changes,shadow changes,pose changes,viewpoint changes and background changes.These complex scene changes pose threat to the robustness and accuracy of the mapping in visual SLAM.In order to meet these challenges,visual SLAM needs loop closure detection,an indispensable module.It can correctly determine whether a mobile robot has visited a certain place based on sensor data,suppress the cumulative errors in localization and mapping and improve the localization accuracy and the mapping quality.The implementation of traditional loop closure detection needs to be based on hand-crafted feature extraction methods.These methods are robust to moderate scene changes,but they cannot cope with complex scene changes.In order to solve this problem,this paper studies how to design efficient and robust loop closure detection algorithms based on the deep learning technologies.The main research results include the following three aspects:(1)Aiming at the problem that existing deep learning based loop closure detection methods fail to design and train the network model from the visual loop closure detection problem itself,which results in the inability to effectively extract low-dimensional discriminative feature descriptors,a multi-constraint deep distance learning method is proposed.Based on this method,a new framework for visual loop closure detection is proposed.First,a convolutional neural network is used to map the original images to the low-dimensional feature space,and the training samples are automatically constructed online.In the training process,a multi-constraint loss function is used to constrain the distance relationship between the feature vectors,thus extracting more discriminative low-dimensional feature descriptors.Compared with the traditional deep distance learning method,this method can achieve the highest accuracy under the same recall rate on the New College and the TUM dataset.(2)Aiming at the problem that existing deep learning based loop closure detection methods still use a pre-defined fixed distance calculation formula and it is difficult to accurately metric the similarity between images,a lightweight relational network architecture is proposed.The architecture uses a data-driven approach to learn feature extraction and similarity metric simultaneously in an end-to-end way,without the need to pre-define the calculation formula for distance metric.At the same time,images before and after frames are constructed into image sequences as input,which can not only provide additional multi-view information,but also accelerate the speed of similarity metric.(3)Aiming at the problems that existing deep learning based loop closure detection methods rely on a large amount of labeled data for supervised training and weak generalization ability,a novel unsupervised loop closure detection method based on binary GAN is proposed.This method uses a novel loss function for spreading the hamming distance relationship and propagates the hamming distance relationship from the high-dimensional feature space to the low-dimensional feature space,thus extracting low-dimensional and discriminative binary feature descriptors.It avoids the time-consuming and labor-intensive data labeling process,and the extracted binary feature descriptors greatly save storage space and computing resources.