Petr Dokládal

Modelling of Overlapping Circular Objects Based on Level Set Approach

The paper focuses on the extraction and modelling of circular objects embedded up to any extent. The proposed method is inspired from the continuous Level Set theory and consists of two stages. First, by using the local curvature and the normal vector on the boundaries are detected the shape parameters of the sought circular objects. Second, an area-based matching procedure detects such cases where several identified circles correspond to only one, partially ocluded object.

A parallel algorithm for solving the Eikonal equation

A large variety of methods based on partial differential equations (PDE) use the interface propagation. For their flexibility these methods are being more and more applied to various problems ranging from physics, fluid mechanics to control theory and computer vision. The solution of the PDE-based interface evolution is in itself a complex iterative computational task involving a great number of iterations (unknown a priori). Therefore, these applications are very demanding on the hardware and their real-time implementation is still a challenging problem. An efficient implementation could be done by using a specific parallel architecture. This paper proposes an original, entirely parallel algorithm to solve the Eikonal equation. Which is the base of applications using a weighted distance function. This algorithm allows the parallel implementation of active contours methods or continuous watershed on a specific hardware.

Embedded real-time architecture for level-set-based active contours

Methods described by partial differential equations have gained a considerable interest because of undoubtful advantages such as an easy mathematical description of the underlying physics phenomena, subpixel precision, isotropy, or direct extension to higher dimensions. Though their implementation within the level set framework offers other interesting advantages, their vast industrial deployment on embedded systems is slowed down by their considerable computational effort. This paper exploits the high parallelization potential of the operators from the level set framework and proposes a scalable, asynchronous, multiprocessor platform suitable for system-on-chip solutions. We concentrate on obtaining real-time execution capabilities. The performance is evaluated on a continuous watershed and an object-tracking application based on a simple gradient-based attraction force driving the active countour. The proposed architecture can be realized on commercially available FPGAs. It is built around general-purpose processor cores, and can run code developed with usual tools.

Contour-based object tracking with gradient-based contour attraction field

Object tracking is a problem which is very often addressed by various approaches, including background subtraction, visual constancy or geometric flow. The contribution of this paper is twofold. First, it proposes an enhanced feature-weighted gradient, enhancing the contours of the object characterized by this feature, and attenuating the gradient amplitude wherever the feature is not found. The second, and main contribution, is a new integrator for active-contours based tracking. It uses the gradient alongside the object contours and generates a bidirectional attraction field. The combination of the contour tracking and the weighted gradient is illustrated on the drivers-face tracking problem, where the illumination conditions (brightness, colour) change as the vehicle passes through different environments.

Asynchronous multi-core architecture for level set methods

The paper proposes an asynchronous multi-core architecture for embedded systems using partial differential equation-based image processing algorithms. A study of data flow and timing analysis is carried out in order to reveal optimal global architecture specifications. The global architecture uses a semi-parallel approach with several processing units running in parallel and shared memory blocks. The results are illustrated by the implementation of a continuous watershed transform, followed by a discussion of the measured execution time and the computational load to demonstrate the efficiency.

MASSIVE MARCHING: A PARALLEL COMPUTATION OF DISTANCE FUNCTION

The methods based on the evolution of a curve controlled by partial differential equations (PDE), represent an efficient and flexible tool of image segmentation, recognition or object tracking. For these methods, an accurate and rapid computation of the distance function is of a key importance. We propose a new, entirely parallel algorithm yielding the distance function and its hardware implementation.