Hélène Laurent

Review and evaluation of commonly-implemented background subtraction algorithms

Locating moving objects in a video sequence is the first step of many computer vision applications. Among the various motion-detection techniques, background subtraction methods are commonly implemented, especially for applications relying on a fixed camera. Since the basic inter-frame difference with global threshold is often a too simplistic method, more elaborate (and often probabilistic) methods have been proposed. These methods often aim at making the detection process more robust to noise, background motion and camera jitter. In this paper, we present commonly-implemented background subtraction algorithms and we evaluate them quantitatively. In order to gauge performances of each method, tests are performed on a wide range of real, synthetic and semi-synthetic video sequences representing different challenges.

Comparative study of background subtraction algorithms

In this paper, we present a comparative study of several state of the art background subtraction methods. Approaches ranging from simple background subtraction with global thresholding to more sophisticated statistical methods have been implemented and tested on different videos with ground truth. The goal of this study is to provide a solid analytic ground to underscore the strengths and weaknesses of the most widely implemented motion detection methods. The methods are compared based on their robustness to different types of video, their memory requirement, and the computational effort they require. The impact of a Markovian prior as well as some post-processing operators are also evaluated. Most of the videos used in this study come from state-of-the-art benchmark databases and represent different challenges such as poor signal-to-noise ratio, multimodal background motion and camera jitter. Overall, this study not only helps better understand to which type of videos each method suits best but also estimate how better sophisticated methods are, compared to basic background subtraction methods.

Unsupervised performance evaluation of image segmentation

We present in this paper a study of unsupervised evaluation criteria that enable the quantification of the quality of an image segmentation result. These evaluation criteria compute some statistics for each region or class in a segmentation result. Such an evaluation criterion can be useful for different applications: the comparison of segmentation results, the automatic choice of the best fitted parameters of a segmentation method for a given image, or the definition of new segmentation methods by optimization. We first present the state of art of unsupervised evaluation, and then, we compare six unsupervised evaluation criteria. For this comparative study, we use a database composed of 8400 synthetic gray-level images segmented in four different ways. Vinets measure (correct classification rate) is used as an objective criterion to compare the behavior of the different criteria. Finally, we present the experimental results on the segmentation evaluation of a few gray-level natural images.

optimization-based image segmentation by genetic algorithms

Many works in the literature focus on the definition of evaluation metrics and criteria that enable to quantify the performance of an image processing algorithm. These evaluation criteria can be used to define new image processing algorithms by optimizing them. In this paper, we propose a general scheme to segment images by a genetic algorithm. The developed method uses an evaluation criterion which quantifies the quality of an image segmentation result. The proposed segmentation method can integrate a local ground truth when it is available in order to set the desired level of precision of the final result. A genetic algorithm is then used in order to determine the best combination of information extracted by the selected criterion. Then, we show that this approach can either be applied for gray-levels or multicomponents images in a supervised context or in an unsupervised one. Last, we show the efficiency of the proposed method through some experimental results on several gray-levels and multicomponents images.

Comparative study of contour detection evaluation criteria based on dissimilarity measures

We present in this article a comparative study of well-known supervised evaluation criteria that enable the quantification of the quality of contour detection algorithms. The tested criteria are often used or combined in the literature to create new ones. Though these criteria are classical ones, none comparison has been made, on a large amount of data, to understand their relative behaviors. The objective of this article is to overcome this lack using large test databases both in a synthetic and a real context allowing a comparison in various situations and application fields and consequently to start a general comparison which could be extended by any person interested in this topic. After a review of the most common criteria used for the quantification of the quality of contour detection algorithms, their respective performances are presented using synthetic segmentation results in order to show their performance relevance face to undersegmentation, oversegmentation, or situations combining these two perturbations. These criteria are then tested on natural images in order to process the diversity of the possible encountered situations. The used databases and the following study can constitute the ground works for any researcher who wants to confront a new criterion face to well-known ones.

Comparison of invariant descriptors for object recognition

This paper deals with the performance evaluation of three object invariant descriptors: Hu moments, Zernike moments and Fourier-Mellin descriptors. Experiments are conducted on a database of 100 objects extracted from the Columbia Object Image Library (COIL-100). Original images from this database only present geometric transformations. They therefore allow to quantify the scale and rotation invariances of the different features and to compare their ability to discriminate objects. In order to test the robustness of the three tested descriptors, we have completed the data set by images including different perturbations: noise, occlusion, luminance variation, backgrounds (uniform, with noise, textured) added to the object. Recognition tests are realised using a support vector machine as supervised classification method. Experimental results are summarized and analyzed permitting to compare the global performances of the different descriptors.

A Real Time Human Detection System Based on Far Infrared Vision

We present in this article a human detection and tracking algorithm using infrared vision in order to have reliable information on a room occupation. We intend to use this information to limit energetic consumption (light, heating). We perform first, a foreground segmentation with a Gaussian background model. A tracking step based on connected components intersections allows to collect information on 2D displacements of moving objects in the image plane. A classification based on a cascade of boosted classifiers is used for the recognition. Experimental results show the efficiency of the proposed algorithm.

A new supervised evaluation criterion for region based segmentation methods

We present in this article a new supervised evaluation criterion that enables the quantification of the quality of region segmentation algorithms. This criterion is compared with seven well-known criteria available in this context. To that end, we test the different methods on natural images by using a subjective evaluation involving different experts from the French community in image processing. Experimental results show the benefit of this new criterion.

Vision-Based System for Human Detection and Tracking in Indoor Environment

In this paper, we propose a vision-based system for human detection and tracking in indoor environment using a static camera. The proposed method is based on object recognition in still images combined with methods using temporal information from the video. Doing that, we improve the performance of the overall system and reduce the task complexity. We first use background subtraction to limit the search space of the classifier. The segmentation is realized by modeling each background pixel by a single Gaussian model. As each connected component detected by the background subtraction potentially corresponds to one person, each blob is independently tracked. The tracking process is based on the analysis of connected components position and interest points tracking. In order to know the nature of various objects that could be present in the scene, we use multiple cascades of boosted classifiers based on Haar-like filters. We also present in this article a wide evaluation of this system based on a large set of videos.

Comparative study of metrics for evaluation of object localisation by bounding boxes.

This paper deals with a comparative study of metrics allowing the evaluation of results provided by object localization algorithms. We particularly focus on localization by the bounding box representation. 26 metrics are studied in this paper. A protocol is presented for the creation of ground truths and synthetic results of localization algorithms. These synthetic results permit to simulate several inaccuracies (translation, scale errors... ) and to study the metrics behaviors face to the considered alterations. Finally, some conclusions and perspectives are given.