SE(2)-Constrained Visual Inertial Fusion for Ground Vehicles

Visual inertial fusion is the problem to estimate vehicle states using visual and inertial measurements, and has recently been an active research topic. Existing visual inertial fusion approaches formulate their estimation algorithms for mobile robots in general 3-D space, and cannot satisfactorily model the nearly planar motion of ground vehicles. This paper proposed a new visual inertial fusion framework for ground vehicles, based on an SE(2)-constrained pose parameterization, which maximally conforms to the realistic situations of the ground vehicle motion. The estimation system is formulated as a graph optimization problem, and the SE(2)-constrained parameterization is implemented by introducing specially designed edges on the vehicle poses in the graph. Keyframe-based optimization and marginalization help to ensure the real-time performance. Experimental tests were conducted on challenging data sets with illumination changes and sharp-turn motion to validate the better accuracy of the proposed system compared with the state-of-the-art visual inertial estimation approaches.