Abstract:

SLAM (Simultaneous Localization And Mapping) broadly used in robots, drones and autonomous cars. However, SLAM estimation accuracy is impacted by sensor noise and featureless environment. The loop detection technique is the key to provide long-term precision. We propose to use the corners geometrical relationship to efficiently encode lidar point cloud scans and combine it with the Bag of Words method to achieve fast and robust loop detection performance. The experiment will be conducted on both large-scale online and self-collected datasets to validate the accuracy and efficiency of our algorithm.