Kshitij will be presenting “Probabilistic Point Cloud Modeling via Self-Organizing Gaussian Mixture Models” at IROS 2023 in Detroit, MI.
When a robot explores an environment, it uses depth and intensity information from sensors like LiDARs and RGB-D cameras to create a perceptual model that must enable obstacle avoidance, reconstruction, and next-best-view selection. Creating these perceptual models at a large scale (spatially and with large numbers of robots) requires processing point cloud data with memory- and communication-efficiency. In our latest pre-print we take a step towards creating such perceptual models. The key idea is to compress point clouds into a 4D continuous probabilistic model. The complexity of this model is adapted with the complexity of the scene.