Odometry-Based Structure from Motion

Structure from motion refers to a technique to obtain 3D information from consecutive images taken with a moving monocular camera. In order to do this, the camera motion performed between two consecutive images needs to be known. In the work reported in this contribution, we investigated the precision of the odometry data of a commercially available passenger car. In order to identify the required precision, we developed an error model based on camera parameters and the bicycle model. We investigated two options, both being based on speed measurements. The first one uses steering angle measurements, the second one uses measurements of the yaw rate. Concluding, we found out that the specified precision of all odometry data available is sufficient to solve structure from motion. Long-term measurements empirically confirm the precision values given in the specification. This result encouraged us to actually implement a structure-from-motion approach which yields depth information as predicted from the theoretical considerations. Further work needs to be carried out in order to compensate for roll motions.