DEPTH IMAGE BASED DIRECT SLAM FOR SMALL UAVS

摘要

We propose a real-time keyframe-based direct(featureless) SLAM method utilizing a singledepth image stream for small UAVs. The realtimeperformance of tracking and mapping isguaranteed on an embedded computer that hasless computational resource. The camera motionis estimated by the direct depth image alignmentbased on keyframes in contrast with previouspoint-cloud approaches. The tracking speed isgreatly improved by the approach and thedecrease in the number of pixels by reduction ofthe image resolution and the intensity gradientbased pixel sampling method. A 3D consistentmap is reconstructed through the pose-graphoptimization. The proposed framework isimplemented on an actual UAV platform with asmall embedded computer, and is compared withthe ground-truth data for validation. It is provento be effective for UAVs that mainly move inlongitudinal direction with rotation.