UAV Visual Teach and Repeat Using Only Semantic Object Features

We demonstrate the use of semantic object detections as robust features for Visual Teach and Repeat (VTR). Recent CNN-based object detectors are able to reliably detect objects of tens or hundreds of categories in a video at frame rates. We show that such detections are repeatable enough to use as landmarks for VTR, without any low-level image features. Since object detections are highly invariant to lighting and surface appearance changes, our VTR can cope with global lighting changes and local movements of the landmark objects. In the teaching phase, we build a series of compact scene descriptors: a list of detected object labels and their image-plane locations. In the repeating phase, we use Seq-SLAM-like relocalization to identify the most similar learned scene, then use a motion control algorithm based on the funnel lane theory to navigate the robot along the previously piloted trajectory. We evaluate the method on a commodity UAV, examining the robustness of the algorithm to new viewpoints, lighting conditions, and movements of landmark objects. The results suggest that semantic object features could be useful due to their invariance to superficial appearance changes compared to low-level image features.