Tracking Visual Landmarks of Opportunity as Rally Points for Unmanned Ground Vehicles

Abstract

In addition to the remote-control for piloting an Unmanned Ground Vehicle (UGV), we aim at adding a new high level command mode: reaching a visual rally point selected by the operator from the camera feed. We modify the recent TransT-M single object tracker to track the rally point as the UGV moves and we couple it with our visual odometry. We design a visibility test to discard false positives, when the rally point disappears from the field of view. We test the tracking on image sequences taken from our platform and the KITTI Vision Benchmark to demonstrate the efficiency and the robustness of the proposed tracking architecture. The visibility test improves the chance of recovery after the landmark disappears and removes false positives. We show that the operator does not need to designate the landmark with high precision as the tracker is tolerant to imprecision on the position and the size of the landmark.