Continuous Thermal VPR in Unmapped Environments

Develop continuous thermal visual place recognition and localization in entirely unmapped environments without visual overlap, addressing the unresolved challenge shared by existing thermal VPR baselines.

Background

The TM-APR framework relies on a fixed reference memory containing coordinate-bearing thermal descriptors and uses online analytic adaptation to refine place-class boundaries. Its spatial gating mechanisms reject corrections when no valid topological reference candidate is available, after which the system falls back to Absolute Pose Regression and motion propagation.

The paper explicitly identifies operation in entirely unmapped regions without visual overlap as an open challenge for thermal VPR systems. In such settings, reference-based loop corrections cannot be applied because the system lacks matching locations in its coordinate memory; the stated unresolved issue concerns achieving continuous navigation under these conditions across thermal VPR baselines.