Enhancing Loop Closure Detection with Object Semantic Scan Context

Loop closure allows a mobile robot to rectify the accumulated drift in its estimated position, facilitating in building an accurate map of the surrounding environment by recognizing previously visited places. 3D point cloud-based loop closure is an essential task for the growing mobile robotics sector.State-of-the-art (SOTA) methods require a robot to revisit a previously explored area in close proximity.To reduce unnecessary robot traversals in the environment, we construct an efficient descriptor, dubbed as Object Semantic Scan Context (OSSC), by encoding semantic features in local descriptors (representations) around external references, i.e., Main Objects (MOs), for accurate loop closure detection even between distant lidar scans. Moreover, we adopt effective strategies for selecting MO(s) and weighting semantic labels to enhance the discriminative power of OSSC in challenging scenarios.Rather than relying on semantic sparsity, OSSC captures the semantic patterns of all objects around MO(s).The proposed descriptor is extensively tested on the \hl{SemanticKITTI and RELLIS-3D} datasets andthe achieved high accuracy in a variety of scenarios, especially with spatially distant scans, corroborate its efficiency and robustness.