Research On Loop Closure Detection In Visual SLAM For Automatic Driving

In automatic driving,SLAM is used for accurate localization and 3D reconstruction.Loop closure detection is helpful for correcting the drift in long run odometry,so that the prediction of car's trajectory can be optimized by it.Loop closure detection methods based on deep learning have been proved to be able to perform better than classical ones in the environment with variable illumination.A research is made in this paper for deeplearning-based loop closure detection from two aspects: background feature extraction and background feature matching.Firstly,an image background feature extraction method under outdoor scenes for automatic driving which is based on background objects detection is proposed.The proposed background feature extraction method is implemented with an object detection net.With a detection of specific background objects in outdoor scenes,the extracted feature is only related to the types and positions of the background objects.Then,a loop closure detection method based on background feature extraction and artificial similarity measurement is proposed.On the basis of the above background feature extraction method,a similarity measurement method based on L2 distance and nearest neighbor search is proposed.Based on this similarity measurement method,background features are extracted from different images in the data set with location information annotation and their similarity scores are calculated to make a match among them.At last,focusing on the similarity measurement and matching between features,another loop closure detection method is proposed,which is based on background feature extraction and metric learning.Proposed method re-extract the background feature to get metric feature through metric learning,which is viewed as the coding feature of image to achieve the final loop closure detection.The performances of two background-feature-extraction-based proposed loop closure detection methods are evaluated and analyzed on MIT-CBCL Street Scenes dataset and KITTI dataset respectively.It is shown in the experimental results that the performance of the two loop detection methods based on background feature extraction is close to the latest research on loop detection algorithm,and has obvious advantages in feature extraction speed.