Computer Science Graduate Seminar

Friday December 3, 2021, 10:00pm

Deep Visual Human Sensing with Application in Robotics



In this talk, I present my thesis contributions to the field of visual human sensing that arise when deploying robots in environments with humans.

After motivating the need for visual human sensing, we start by describing a novel human detector based on a 2D lidar sensor (e.g. a "laser scanner"). It is the first of its kind that is learning-based and general, specifically it does not encode a "two leg prior".

Detection being covered, we move on to discuss person re-identification, and specifically our contribution of establishing triplet-loss based methods as a strong contender and principled approach in the field. Using this we also sketch the way to a completely novel approach on tracking which leverages such triplet-based re-identification models at its core.

We then discuss more detailed analysis of individual persons, specifically their head orientation, which can serve as a cue for their intent or an indicator of what is interesting in the scene, among other things. We derive a novel cyclic regression loss based on the von-Mises distribution and use it, coupled with our "biternion" output layer, to learn continuous regression models using only discrete, weakly labeled data.

Finally, we present a holistic system integrating all of these pieces and several more, highlighting the system-level difficulties of such integration, and proposing some ways around them.


The computer science lecturers invite interested people to join.