EM-Fusion: Dynamic Object-Level SLAM With Probabilistic Data Association

Abstract

The majority of approaches for acquiring dense 3D environment maps with RGB-D\ncameras assumes static environments or rejects moving objects as outliers. The\nrepresentation and tracking of moving objects, however, has significant\npotential for applications in robotics or augmented reality. In this paper, we\npropose a novel approach to dynamic SLAM with dense object-level\nrepresentations. We represent rigid objects in local volumetric signed distance\nfunction (SDF) maps, and formulate multi-object tracking as direct alignment of\nRGB-D images with the SDF representations. Our main novelty is a probabilistic\nformulation which naturally leads to strategies for data association and\nocclusion handling. We analyze our approach in experiments and demonstrate that\nour approach compares favorably with the state-of-the-art methods in terms of\nrobustness and accuracy.\n