Optimal motion estimation from visual and inertial measurements

Abstract

Abstract

Efficient Online Egomotion Estimation Using Visual and Inertial Readings

“…It can be argued and past research has established that combining visual and inertial sensor readings to determine egomotion yields much better results than those obtained using only one type of reading. This result is both intuitive and has been empirically demonstrated repeatedly [16] [17] [15] [1]. It seems that cameras and inertial measurement units (IMU) complement each other very well : for discontinuous motions where the visual feature tracker fails to make sense of incoming images, the IMU excels, whereas in slower smoother motions where the IMU's readings are drowned in noise, the visual feature tracker performs well.…”

“…Furthermore, to establish a basis for comparison with other work, we reproduce the main test from [16] and [17] which claim to produce optimal egomotion estimates via batch and online algorithms. The results presented below were obtained using 17 features with varying depths and spanning the 4 quadrants of image space.…”

“…For instance, [15] and [1] use Kalman filters (or Kalman filter extensions) to integrate measurements produced by both their measurement devices whereas [16] and [17] use Kalman filters to estimate egomotion, inertial readings, estimation errors and the 3D coordinates of all known features. The latter approach seems somewhat out of place in the context of this work, i.e.…”

“…The main difference is that this test demonstrates proper tracking of rotations rather than translations. "Clover Leaf" Test [16] [17]: This test meant to compare our approach's performance with that of the approach from [16] and [17] which claims to produce optimal egomotion estimates via batch and online algorithms. Hence, this test described a motion very similar to the one used in the evaluations of those works.…”

“…Several authors make use of statistical filters to fuse inertial and visual readings [16] [17] [15] [1], many of which produce noteworthy results. For instance, [15] and [1] use Kalman filters (or Kalman filter extensions) to integrate measurements produced by both their measurement devices whereas [16] and [17] use Kalman filters to estimate egomotion, inertial readings, estimation errors and the 3D coordinates of all known features.…”