Localizing using 3D point clouds has been a long-standing challenge for many years. LiDAR sensors are capable of collecting 3D point clouds, from which egocentric global descriptors can be generated to recognize previously visited locations. One of the most well-known  global descriptors is Scan Context, which has inspired numerous similar methods over the years. However, these descriptors are typically two-dimensional and require column offsets for distance calculation, resulting in high computational costs. To overcome this, many methods use simplified one-dimensional descriptors (such as RingKey) and multi-phase searches for localization. In this research, we explored the possibility of binarizing global descriptors. We applied established binarization techniques and evaluated their performance, storage requirements, and runtime. We also proposed two new binarization methods specifically for egocentric global descriptors. Our results demonstrate that the best binarization methods achieve performance comparable to the original descriptors, while reducing storage requirements and runtime by ten times.

Eredeti fellelhetőség: dl.acm.org