BALM: Bundle Adjustment for Lidar Mapping

A local Bundle Adjustment (BA) on a sliding window of keyframes has been widely used in visual SLAM and proved to be very effective lowering the drift. But lidar SLAM, BA method is hardly because sparse feature points (e.g., edge plane) make exact point matching impossible. In this letter, we formulate as minimizing distance from its matched or plane. Unlike (and prior plane adjustment SLAM) where co-determined along with pose, show that can analytically solved removed BA, resultant only dependent scan poses. This greatly reduces optimization scale allows large-scale dense features used. To speedup optimization, derive analytical derivatives cost function, up second order, closed form. Moreover, propose novel adaptive voxelization search correspondence efficiently. The proposed formulations are incorporated into LOAM back-end for map refinement. Results that, although back-end, efficiently, even real-time at 10 Hz when optimizing 20 scans point-cloud. also considerably lowers Our implementation open-sourced benefit community. 1 https://github.com/hku-mars/BALM.