LAUSR

With growing interest in visual/inertial state estimation and an increasing number of approaches and applications emerging for this technology, camera/IMU calibration can be a valuable tool to increase the performance of these methods and to further the understanding of the involved sensor modalities. In this paper, we assess the impact of two different adjustments to the commonly used sensor models. First, we extend the IMU model to take the displacement of individual accelerometer axes into account. We show that especially high quality devices benefit from this extension, since these IMUs often employ separate sensors for each axis. Second, we propose a novel, direct model for the camera measurements that operates on image intensities rather than corner positions. This formulation is capable of explicitly accounting for motion blur and defocus, but it requires significant modeling efforts. Our results demonstrate that the transformation between camera and IMU can be estimated to a precision exceeding $\tfrac {1}{5}\mathrm { \text {m} \text {m} }$ and $\tfrac {1}{100}\mathrm { ^{\circ}}$ , while temporal offsets are determined to microsecond precision—on data sets of merely 20-s length. At the same time, image exposure time can be inferred to an accuracy of about $\tfrac {2}{100}\mathrm { \text {m} \text {s} }$ from motion blur.