A comparison of linearisation and the unscented transform for computer vision applications

Accurate sensor noise propagation is critical for many computer vision and robotic applications. Several probabilistic computer vision techniques require estimates of sensor noise after it has been propagated through one or many nonlinear transformations. We investigate the unscented transform as an alternative to the standard linearisation technique for uncertainty propagation in a computer vision framework. An evaluation is performed using synthetic data for two common computer vision sensors, an RGB-D sensor and stereo camera pair. The unscented transform is shown to outperform linearisation when used to estimate distributions of reconstructed, 3D points from image features. Experimental results also indicate that the unscented transform is a viable replacement for linearisation when used in a probabilistic visual odometry framework.