Felix Woelk

Calibration of a multi-camera rig from non-overlapping views

A simple, stable and generic approach for estimation of relative positions and orientations of multiple rigidly coupled cameras is presented in this paper. The algorithm does not impose constraints on the field of view of the cameras and works even in the extreme case when the sequences from the different cameras are totally disjoint (i.e. when no part of the scene is captured by more than one camera). The influence of the rig motion on the existence of a unique solution is investigated and degenerate rig motions are identified. Each camera captures an individual sequence which is afterwards processed by a structure and motion (SAM) algorithm resulting in positions and orientations for each camera. The unknown relative transformations between the rigidly coupled cameras are estimated utilizing the rigidity constraint of the rig.

Extraction of 3D freeform surfaces as visual landmarks for real-time tracking

This work presents a system for the generation of a free-form surface model from video sequences. Although any single centered camera can be applied in the proposed system the approach is demonstrated using fish-eye lenses because of their good properties for tracking. The system is designed to function automatically and to be flexible with respect to size and shape of the reconstructed scene. To minimize geometric assumptions a statistic fusion of dense depth maps is utilized. Special attention is payed to the necessary rectification of the spherical images and the resulting iso-disparity surfaces, which can be exploited in the fusion approach. Before dense depth estimation can be performed the cameras’ pose parameters are extracted by means of a structure-from-motion (SfM) scheme. In this respect automation of the system is achieved by thorough decision model based on robust statistics and error propagation of projective measurement uncertainties. This leads to a scene-independent set of only a few parameters. All system components are formulated in a general way, making it possible to cope with any single centered projection model, in particular with spherical cameras. In using wide field-of-view cameras the presented system is able to reliably retrieve poses and consistently reconstruct large scenes. A textured triangle mesh constructed on basis of the scene’s reconstructed depth, makes the system’s results suitable to function as reference models in a GPU driven analysis-by-synthesis framework for real-time tracking.

Fast Monocular Bayesian Detection of Independently Moving Objects by a Moving Observer

A fast algorithm for the detection of independently moving objects by an also moving observer by means of investigating optical flow fields is presented. Since the measurement of optical flow is a computationally expensive operation, it is necessary to restrict the number of flow measurements. The proposed algorithm uses two different ways to determine the positions, where optical flow is calculated. A part of the positions is determined using a particle filter, while the other part of the positions is determined using a random variable, which is distributed according to an initialization distribution. This approach results in a restricted number of optical flow calculations leading to a robust real time detection of independently moving objects on standard consumer PCs.

A monocular image based intersection assistant

Since cross traffic in inner city intersections is potentially dangerous, its early detection is desirable. In this paper an image-based driver-assistant system is presented. The investigation of optical flow, computed from an image sequence, recorded by a single pan-tilt camera, is used as basis to this system. Its goal is to direct the drivers attention to moving objects in inner city intersections.

Robust monocular detection of independent motion by a moving observer

A fast and robust algorithm for the detection of independently moving objects by a moving observer by means of investigating optical flow fields is presented. The detection method for independent motion relies on knowledge about the camera motion. Even though inertial sensors provide information about the camera motion, the sensor data does not always satisfy the requirements of the proposed detection method. The first part of this paper therefore deals with the enhancement of earlier work [29] by ego-motion refinement. A linearization of the ego-motion estimation problem is presented. Further on a robust enhancement to this approach is given. Since the measurement of optical flow is a computationally expensive operation, it is necessary to restrict the number of flow measurements. The proposed algorithm uses two different ways to determine the positions, where optical flow is calculated. A fraction of the positions is determined by using a sequential Monte Carlo sampling resampling algorithm, while the remaining fraction of the positions is determined by using a random variable, which is distributed according to an initialization distribution. This approach results in a fixed number of optical flow calculations leading to a robust real time detection of independently moving objects on standard consumer PCs.

A monocular collision warning system

A system for the detection of independently moving objects by a moving observer by means of investigating optical flow fields is presented. The usability of the algorithm is shown by a collision detection application. Since the measurement of optical flow is a computationally expensive operation, it is necessary to restrict the number of flow measurements. The first part of the paper describes the usage of a particle filter for the determination of positions where optical flow is calculated. This approach results in a fixed number of optical flow calculations leading to a robust real time detection of independently moving objects on standard consumer PCs. The detection method for independent motion relies on knowledge about the camera motion. Even though inertial sensors provide information about the camera motion, the sensor data does not always satisfy the requirements of the proposed detection method. The second part of this paper therefore deals with the enhancement of the camera motion using image information. The third part of this work specifies the final decision module of the algorithm. It derives a decision (whether to issue a warning or not) from the sparse detection information.

An airborne Bayesian color tracking system

Mobile tracking systems are of wide interest. For instance sport photographers often hire helicopters to obtain close range images of e.g. aquatic athletes. Besides the high price for hiring a helicopter, good helicopter pilots are not everywhere available. One way to circumvent these problems is the use of a cheap model helicopter. Since flying a model helicopter is a demanding task, there are either two persons needed for the operation of the system (one person for steering the helicopter and the other for the operation of the camera), or alternatively an automatic tracking system can be used. In this work an automatic, image based tracking system is described. The algorithm combines color histograms as similarity measures with a particle filter and achieves fast and robust tracking results. A novel method for fast histogram computation is proposed. The system keeps the object of interest automatically in the viewing field of the camera so that the only task left to the camera operator is to initiate the imaging process. This simple task can be additionally managed by the helicopter pilot so that a single person is sufficient to steer the helicopter and to take the images.