We present a framework that allows for localization based on very low resolution omnidirectional image data using regression techniques. Previous related methods are constrained to image data labeled with exact position information acquired in the training phase. We relax this constraint and propose to learn local heteroscedastic Gaussian processes by accumulating odometry data which can easily be acquired. The processes are used as a probabilistic map to predict recording positions of newly acquired images by a fusion of the uncertain training data. In contrast to many feature-based approaches, our framework does not rely on any explicit correspondences over images as well as over positions and only imposes very weak assumptions on the type and quality of the image representations.

We investigate the application of particle filters to estimate the orientation of a mobile agent based on an omnidirectional video stream. By applying spherical signal analysis to sequences of low resolution input images we perform a real-time estimation of the relative camera rotation. We use normalized cross-correlation computed with a fast frequency domain approach that yields unbiased estimates which can be further processed by a particle filter. We discuss the quaternion representation of the rotational state space and evaluate methods for meaningful averaging. A prototype system is presented and experiments with real image sequences show that robust estimation of the rotational motion is achievable even when the input images are corrupted, e.g. due to occlusions.

We present a probabilistic localization and orientation estimation method for mobile agents equipped with omnidirectional vision. In our appearance-based framework, a scene is learned in an offline step by modeling the variation of the image energy in the frequency domain via Gaussian process regression. The metric localization of novel views is then solved by maximizing the joint predictive probability of the Gaussian processes using a particle filter which allows to incorporate a motion model in the prediction step. Based on the position estimate, a synthetic view is generated and used as a reference for the orientation estimation which is also performed in the Fourier space. Using real as well as virtual data, we show that this framework allows for robust localization in 2D and 3D scenes based on very low resolution images and with competitive computational load.