Ambient motion estimation in dynamic scenes using wearable visual-inertial sensors

This paper proposes a method to estimate the motion of ambient objects including translational and rotational velocities by moving observers with hybrid visual-inertial sensors. Ambient motion is recovered from visual optical flows that represent ego and ambient dynamics. In this paper, each moving object is considered as a rigid body that has been segmented from the background using computer vision algorithms. In motion recovery, the fundamental challenge is to resolve the coupling between scene depths and translational velocities. Ambient rotational velocities are obtained following a depth-independent bilinear constrain. The scales of ambient trans-lational velocities is computed using the proposed dynamics constraint with an assumption that ambient accelerations are negligible. A fix-point optimization scheme is further introduced to iteratively refine the recoveries of translational and rotational ambient motion until an expected precision is achieved or the maximal iteration is reached. During the optimization, translational ambient motion is precisely recovered and translational ambient motion is rescaled to the canonical amplitude. The results of the simulation study show the effectiveness of the proposed method in motion analysis and prediction.