TwistSLAM: Constrained SLAM in Dynamic Environment

Classical visual simultaneous localization and mapping (SLAM) algorithms usually assume the environment to be rigid. This assumption limits applicability of those as they are unable accurately estimate camera poses world structure in real life scenes containing moving objects (e.g. cars, bikes, pedestrians, etc.). To tackle this issue, we propose TwistSLAM: a semantic, dynamic stereo SLAM system that can track environment. Our algorithm creates clusters points according their semantic class. Thanks definition inter-cluster constraints modeled by mechanical joints (function class), novel constrained bundle adjustment is then able jointly both velocities along with classical trajectory. We evaluate our approach on several sequences from public KITTI dataset demonstrate quantitatively it improves object tracking compared state-of-the-art approaches.