Alain Crouzil

A new correlation criterion based on gradient fields similarity

Correlation based methods are a common tool for the correspondence problem. They are able to perform a dense point-to-point matching between two images, but post-processing is often necessary to improve the results. In this paper we propose a new correlation measure using the gradient vector fields of the images. We compare our method to classical correlation measures based on the grey levels. The new method gives better results than others when it is applied on random dot stereograms and on real images.

A multiresolution approach for shape from shading coupling deterministic and stochastic optimization

Shape from shading is an ill-posed inverse problem for which there is no completely satisfactory solution in the existing literature. In this paper, we address shape from shading as an energy minimization problem. We first show that the deterministic approach provides efficient algorithms in terms of CPU time, but reaches its limits since the energy associated with shape from shading can contain multiple deep local minima. We derive an alternative stochastic approach using simulated annealing. The obtained results strongly outperform the results of the deterministic approach. The shortcoming is an extreme slowness of the optimization. Therefore, we propose a hybrid approach which combines the deterministic and stochastic approaches in a multiresolution framework.

Towards shape from shading under realistic photographic conditions

This paper describes a new modeling of the shape from shading problem taking perspective projection into account, and proposes a method of resolution for the new equation. An application is proposed, which consists in correcting the defects of photographs of skew i.e., nonflat, documents.

Similarity measures for image matching despite occlusions in stereo vision

In the context of computer vision, matching can be done with similarity measures. This paper presents the classification of these measures into five families. In addition, 18 measures based on robust statistics, previously proposed in order to deal with the problem of occlusions, are studied and compared to the state of the art. A new evaluation protocol and new analyses are proposed and the results highlight the most efficient measures, first, near occlusions, the smooth median powered deviation, and second, near discontinuities, a non-parametric transform-based measure, CENSUS.

Towards correlation-based matching algorithms that are robust near occlusions

In the context of computer vision, matching can be done using correlation measures. This paper presents new algorithms that use two correlation measures: the zero mean normalised cross-correlation, ZNCC, and the smooth median absolute deviation, SMAD. While ZNCC is efficient in non-occluded areas and non-robust near occlusions, SMAD is non-efficient in non-occluded areas and robust near occlusions. The aim is to use the advantages of ZNCC and SMAD to deal with the problem of occlusions and to obtain dense disparity maps. The experimental results show that these algorithms are better than the ZNCC-based algorithm and SMAD-based algorithm.

On Constant Focal Length Self-Calibration From Multiple Views

We investigate the problem of finding the metric structure of a general 3D scene viewed by a moving camera with square pixels and constant unknown focal length. While the problem has a concise and well-understood formulation in the stratified framework thanks to the absolute dual quadric, two open issues remain. The first issue concerns the generic Critical Motion Sequences, i.e. camera motions for which self-calibration is ambiguous. Most of the previous work focuses on the varying focal length case. We provide a thorough study of the constant focal length case. The second issue is to solve the nonlinear set of equations in four unknowns arising from the dual quadric formulation. Most of the previous work either does local nonlinear optimization, thereby requiring an initial solution, or linearizes the problem, which introduces artificial degeneracies, most of which likely to arise in practice. We use interval analysis to solve this problem. The resulting algorithm is guaranteed to find the solution and is not subject to artificial degeneracies. Directly using interval analysis usually results in computationally expensive algorithms. We propose a carefully chosen set of inclusion functions, making it possible to find the solution within few seconds. Comparisons of the proposed algorithm with existing ones are reported for simulated and real data.

Colour Correlation-based Matching

In the context of computer vision, stereo matching can be done using correlation measures. Few papers deal with color correlationbased matching so the underlying problem of this paper is about how it can be adapted to color images. The goals of this work are to help choosing a color space and to generalize the correlation measures to color. Nine color spaces and three different methods have been investigated to evaluate their suitability for stereo matching. The results show us to what extent stereo matching can be improved with color.

Towards a guaranteed solution to plane-based self-calibration

We investigate the problem of self-calibrating a camera, from multiple views of a planar scene. By self-calibrating, we refer to the problem of simultaneously estimate the camera intrinsic parameters and the Euclidean structure of one 3D plane. A solution is usually obtained by solving a non-linear system via local optimization, with the critical issue of parameter initialization, especially the focal length. Arguing that these five parameters are inter-dependent, we propose an alternate problem formulation, with only three d.o.f., corresponding to three parameters to estimate. In the light of this, we are concerned with global optimization in order to get a guaranteed solution, with the shortest response time. Interval analysis provides an efficient numerical framework, that reveals to be highly performant, with regard to both estimation accuracy and time-consuming.

A region-based randomized voting scheme for stereo matching

This paper presents a region-based stereo matching algorithm which uses a new method to select the final disparity: a random process computes for each pixel different approximations of its disparity relying on a surface model with different image segmentations and each found disparity gets a vote. At last, the final disparity is selected by estimating the mode of a density function built from these votes. We also advise how to choose the different parameters. Finally, an evaluation shows that the proposed method is efficient even at sub-pixel accuracy and is competitive with the state of the art.

Automatic vehicle counting system for traffic monitoring

Abstract. The article is dedicated to the presentation of a vision-based system for road vehicle counting and classification. The system is able to achieve counting with a very good accuracy even in difficult scenarios linked to occlusions and/or presence of shadows. The principle of the system is to use already installed cameras in road networks without any additional calibration procedure. We propose a robust segmentation algorithm that detects foreground pixels corresponding to moving vehicles. First, the approach models each pixel of the background with an adaptive Gaussian distribution. This model is coupled with a motion detection procedure, which allows correctly location of moving vehicles in space and time. The nature of trials carried out, including peak periods and various vehicle types, leads to an increase of occlusions between cars and between cars and trucks. A specific method for severe occlusion detection, based on the notion of solidity, has been carried out and tested. Furthermore, the method developed in this work is capable of managing shadows with high resolution. The related algorithm has been tested and compared to a classical method. Experimental results based on four large datasets show that our method can count and classify vehicles in real time with a high level of performance (>98%) under different environmental situations, thus performing better than the conventional inductive loop detectors.