A 2D-Laser-based Pedestrian Detector for Mobile Robots via Self-Supervised Transfer Learning

The applications of mobile robots have gradually expanded to complex environments in human daily life. These environments often involve pedestrian activities that possess diverse purposes and statuses. Consequently, it becomes necessary to introduce pedestrian detection models that can predict the exact locations of pedestrians. However, the efficacy of 2D-Laser-based pedestrian detection models trained on open-source datasets is hindered by the difficulty of deploying them on mobile robots. The laser point cloud data from the mobile robot differs in device and resolution. Therefore, it’s non-trivial to deploy a pedestrian detection model trained on open-source datasets to a physical robot. To tackle this problem, we propose a pedestrian detection framework based on self-supervision. The framework is comprised of a pre-trained laser-based detector and a self-supervised data annotator. Our proposed method allows for rapid transfer learning by aligning the laser and camera. The data annotator employs instance segmentation on camera images to generate pseudo-labels for laser point clouds. These pseudo-labels are then used to bridge the gap between open-source data and laser point cloud data collected from mobile robots through transfer learning. Extensive experiments and visualizations reveal the efficacy of the proposed framework.