Frank Ade

Tracking multiple objects using the Condensation algorithm

Abstract Some years ago a new tracker, the Condensation algorithm, came to be known in the computer vision community. It describes a stochastic approach that has neither restrictions on the system and measurement models used nor on the distributions of the error sources, but it cannot track an arbitrary changing number of objects. In this paper an extension of the Condensation algorithm is introduced that relies on a single probability distribution to describe the likely states of multiple objects. By introducing an initialization density, observations can flow directly into the tracking process, such that newly appearing objects can be handled.

Feature extraction and decision procedure for automated inspection of textured materials

This paper proposes a general system approach applicable to the automatic inspection of textured material. First, the input image is preprocessed in order to be independent of non-uniformities. A tone-to-texture transform is then performed by mapping the original grey level picture on a multivariate local feature sequence, which turns out to be normally distributed. More specifically, features derived with the help of the Karhunen-Loeve decomposition of a small neighbourhood of each pixel are used. A decision as to conformity with a reference texture is arrived at by thresholding the Mahalanobis distance for every realization of the feature vector. It is shown that this approach is optimum under the Gaussian assumption in the sense that it has a minimum acceptance region for a fixed probability of false rejection.

Grouping Symmetrical Structures for Object Segmentation and Description

A method is presented for segmenting gray-value images into objects (or their parts) and for recognizing the detected objects. Starting from edge maps, the method extracts axial descriptions of symmetrical shapes. Initially, a piecewise linear approximation of the binary edge map is obtained. From any two of the resulting linear segments, a Linear Segment Pair (LSP) is formed and several of its attributes are computed. These attributes allow the method to reject or select the LSPs through symbolic rules and coarse numeric thresholds. Grouping the LSPs into couples is governed by additional attributes and rules, with the final representation consisting of ordered sets of LSPs. The application to shape description, object recognition, and stereo correspondence is presented. This segmentation method is useful for a broad range of images; it has been used in a robot vision system which is capable of manipulating three-dimensional, overlapping, real-world objects in close to real time.

Tracking cars in range images using the CONDENSATION algorithm

The detection of objects in every frame of a sequence is often not sufficient for scene interpretation. Tracking can increase the robustness, especially when occlusions occur or when objects temporarily disappear. In this paper we present a stochastic tracking approach which is based on the CONDENSATION algorithm (conditional density propagation over time) that is capable of tracking multiple objects with multiple hypotheses in range images. A probability density function describing the likely state of the objects is propagated over time using a dynamic model. The measurements influence the probability function and allow the incorporation of new objects into the tracking scheme. Additionally, the representation of the density function with a fixed number of samples ensures a constant running time per iteration step. Results with data from different sources are shown for automotive applications.

Tracking cars in range image sequences

Cars will eventually be equipped with a control system to either warn the driver when he gets too close to another car, or even to keep him automatically at an adequate distance. In both cases automatic traffic scene analysis will be needed. We want to base this on a new optical range sensor which acquires distance and intensity information. We present a detection and tracking scheme which is fast, simple and operates with a small basic instruction set. Our region growing algorithm ensures robust segmentation and detects objects even in ambiguous data. Objects are represented by bounding boxes and tracked along the image sequence. The tracking process also provides the velocities for calculating the time to collision with each object. We show results of some experiments with simulated traffic scenes.

Towards Smarter Cars

Most approaches for vision systems use greyscale or color images. In many applications, such as driver assistance or presence detection systems, the geometry of the scene is more relevant than the reflected brightness information and therefore range sensors are of increasing interest. In this paper we focus on an automotive application of such a range camera to increase safety on motorways. This driver assistance system is capable of automatically keeping the car at an adequate distance or warning the driver in case of dangerous situations. The problem is addressed in two steps: obstacle detection and tracking. For obstacle detection two different approaches are presented based on slope evaluation and computation of a road model. For tracking, one approach applies a matching scheme, the other uses a Kalman filter. Results are shown for several experiments.

A rule-based system for house reconstruction from aerial images

We present a system for the reconstruction of houses from aerial images. Most of the methods proposed so far use rather specific models. These models do not hold when dealing with Europeanised houses. Our system works with rules instead of models merely based on shape. This enables us to include additional domain specific knowledge, allowing a larger variability of the objects. However this also necessitates a more complex structure of the knowledge base and a more sophisticated reasoning control. We explain the practical application of such a system for the reconstruction of a house from an aerial image.

The Role of Artificial Intelligence in the Reconstruction of Man-made Objects from Aerial Images

A number of different definitions of artificial intelligence is considered and their possible significance for the domain of reconstruction of man-made objects from aerial images is discussed. A useful property of AI is its inclination towards exploratory work which has helped to launch many new research areas. It is shown that contributions of AI can be found in various subdomains which are important for the field of man-made object reconstruction as a whole, namely perceptual organization, modeling and knowledge representation, control, handling of uncertainties, to name but the most important ones. As the field is far from mature at this very moment, AI can and will further contribute to its advancement.

Segmentation and Symbolic Description of Range Images

This paper describes a method to segment a range image of a scene containing simple objects and to generate a first symbolic description thereof. The most important underlying assumption is that of surface coherence, i.e. that the visible surfaces which make up the objects are piecewise smooth. Each such piece can thus be approximated to any desired degree by analytic functions, e.g. polynomials. The first step consists of clustering of surface normals through the iterative detection of peaks in histograms of surface normal components. The amount of tolerated deviation of normal directions in a cluster is made dependent on the noise level in the range image. The result is a set of patches which correspond to true planar surfaces or to small pieces of curved surfaces. A first region growing recovers points which were discarded because of large normal deviation but which are near to the fitting plane. Now a second order polynomial fit is computed for all patches. Curvature values for the mass centers of the patches are obtained from the polynomial fit surface and are used to recognize the surface type. Further region growing is now performed. Region merging based on surface type and numeric values of attributes is then done. Each region is described and relations between adjacent regions are assembled into an attributed graph. This graph will be used in object recognition.

Vision for a 3-D Object Manipulation System

This contribution describes the vision module of a 3-D object manipulation system, which was implemented in the framework of the so-called COR project1. This large interdisciplinary project was promoted by the Mechatronics working group of the Swiss Federal Institute of Technology. The Mechatronics group comprises the Institutes of Robotics, Electronics, Control Theory, Communication Technology and the Chair for Electrotechnical Constructions.