| Loop closure detection is an important component in a SLAM system.It enables the robot to recognize a place it visited before,which eliminates cumulative error and maintains the consistency of the map.Loop closure detection is an image matching problem.Traditional algorithms are based on the "Bag of Words" method in which feature points are utilized.However,many indoor scenes lack of point features,thus leading to a performance degradation of loop closure.On the other hand,lines in indoor environments often carry much information such as structure and direction,which makes them useful in indoor navigation.As a result,a robust and effective line descriptor is very necessary.For the above two issues,we propose a novel binary line descriptor called BRLD,and implement a loop closure detection system combined with point and line features.We first introduce the background and related work of SLAM and loop closure detection.Then we introduce the BoW method and the parts of feature extraction and description in it.Based on the related work of feature descriptor.After that,we propose a binary line descriptor called BRLD and its computational method.To enhance the robustness of the system built on the basic BoW frame,we also introduce the consistency check and probabilistic model.In our experiments,we discuss the choice of training sets and parameters in BRLD,and test its robustness.We compare BRLD with the state-of-art LBD descriptor.The results show that BRLD's matching performance is close to LBD's,but ten times faster in both descriptor composition and matching.We analyze the performance of visual words generated by feature points and lines.Since each of them has its own advantage,we implement a loop closure detection system based on combining point and line features.Experimental results show that with additional consistency check and probabilistic model,our hybrid loop-closure detection algorithm can achieve almost 100% precision while maintaining a high recall rate on our dataset. |
PDF Full Text Request