Catálogo de publicaciones - libros

This paper proposes a fast 3D head tracking method that is working robustly under a variety of difficult conditions such as the rapidly changing pose, head movement and illumination. First, we obtain the pose robustness by using the 3D cylindrical head model (CHM) and dynamic template. Second, we also obtain the robustness about the fast head movement by using the dynamic template. Third, we obtain the illumination robustness by modeling the illumination basis vectors and by adding them to the previous input image to adapt the current input image. Additionally, to make it more robust, we use a re-registration technique that takes the stored reference image as the template when the registration error becomes great. Experimental results show that the proposed head tracking method outperforms the other tracking method using the fixed and dynamic template in terms of the small pose error and the higher successful tracking rate and it tracks the head successfully even if the head moves fast under the rapidly changing poses and illuminations in a speed of 10-15 frames/sec.

This paper addresses the problem of detecting and tracking multiple moving people when the scene background is not known in advance. We have proposed a new background detection technique for dynamic environment that learns and models the scene background based on K-mean clustering technique and pixel statistics. The background detection is achieved using the first frames of the scene where, the number of these frames needed depends on how dynamic is the observed environment. We have also proposed a new feature-based framework, which requires feature extraction and feature matching, for tracking moving people. We have considered color, size, blob bounding box and motion information as features of people. In our feature-based tracking system, we have used Pearson correlation coefficient for matching feature-vector with temporal templates. The occlusion problem has been solved by sub-blobbing. Our tracking system is fast and free from assumptions about human structure. The tracking system has been implemented using Visual C++ and OpenCV and tested on real-world videos. Experimental results suggest that our tracking system achieved good accuracy and can process videos close to real-time.

While existing video object tracking is sensitive to the accuracy of object segmentation, we propose a central point based algorithm in this paper to allow inaccurately segmented objects to be tracked inside video sequences. Since object segmentation remains to be a challenge without any robust solution to date, we apply a region-grow technique to further divide the initially segmented object into regions, and then extract a central point within each region. A macro-block is formulated via the extracted central point, and the object tracking is carried out through such centralized macroblocks and their directional vectors. As the size of the macroblock is often much smaller than the segmented object region, the proposed algorithm is tolerant to the inaccuracy of object segmentation. Experiments carried out show that the proposed algorithm works well in tracking video objects measured by both efficiency and effectiveness.

We propose visual tracking of multiple objects (faces of people) in a meeting scenario based on low-level features such as skin-color, target motion, and target size. Based on these features automatic initialization and termination of objects is performed. Furthermore, on-line learning is used to incrementally update the models of the tracked objects to reflect the appearance changes. For tracking a particle filter is incorporated to propagate sample distributions over time. We discuss the close relationship between our implemented tracker based on particle filters and genetic algorithms. Numerous experiments on meeting data demonstrate the capabilities of our tracking approach.

In spite of inaccurate segmentation, the performance of region-based image retrieval is still restricted by the diverse appearances of semantic-similar objects. On the contrary, humans’ linguistic description of image objects can reveal object information at a higher level. Using partial annotated region collection as “visual dictionary”, this paper proposes a semantic sensitive region retrieval framework using middle-level visual & textual object description. To achieve this goal, firstly, a partial image database is segmented into regions, which are manually annotated by keywords to construct a visual dictionary. Secondly, to associate appearance-diverse, semantic-similar objects together, a Bayesian reasoning approach is adopted to calculate the semantic similarity between two regions. This inference method utilizes the visual dictionary to bridge un-annotated images region together at semantic level. Based on this reasoning framework, both query-by-example and query-by-keyword user interfaces are provided to facilitate user query. Experimental comparisons of our method over Visual-only region matching method indicate its effectiveness in enhancing the performance of region retrieval and bridging the semantic gap.

This paper presents a novel surveillance video indexing and retrieval system based on object features similarity measurement. The system firstly extracts moving objects from the videos by an efficient motion segmentation method. The fundamental features of each moving object are then extracted and indexed into the database. During retrieval, the system matches the query with the features indexed in the database without re-processing the videos. Video clips which contain the objects with sufficiently high relevance scores are then returned. The novelty of the system includes: 1. A real-time automatic indexing methodology achieved by a fast motion segmentation, such that the system is able to perform on-the-fly indexing on video sources; and 2. an object-based retrieval system with fundamental features matching approach, which allows user to specify the query by providing an example image or even a sketch of the desired objects. Such an approach can search the desired video clips in a more convenient and unambiguous way comparing with traditional text-based matching.

This paper introduces a composite relevance feedback approach for image retrieval using transaction-based and SVM-based learning. A transaction repository is dynamically constructed by applying these two learning techniques on positive and negative session-term feedback. This repository semantically relates each database image to the query images having been used to date. The query semantic feature vector can then be computed using the current feedback and the semantic values in the repository. The correlation measures the semantic similarity between the query image and each database image. Furthermore, the SVM is applied on the session-term feedback to learn the hyperplane for measuring the visual similarity between the query image and each database image. These two similarity measures are normalized and combined to return the retrieved images. Our extensive experimental results show that the proposed approach offers average retrieval precision as high as 88.59% after three iterations. Comprehensive comparisons with peer systems reveal that our system yields the highest retrieval accuracy after two iterations.

Shape-based image retrieval is one of the most challenging aspects in Content-Based Image Retrieval (CBIR). A variety of techniques are reported in the literature that aim to retrieve objects based on their shapes; each of these techniques has its advantages and disadvantages. In this paper, we propose a novel scheme that exploits complementary benefits of several shape-based image retrieval techniques and integrates their assessments based on a pair-wise co-ranking process. The proposed scheme can handle any number of CBIR techniques; however, three common techniques are used in this study: Invariant Zernike Moments (IZM), Multi-Triangular Area Representation (MTAR), and Fourier Descriptor (FD). The performance of the proposed scheme is compared with that of each of the selected techniques. As will be demonstrated in this paper, the proposed co-ranking scheme exhibits superior performance.

With the explosive growth of the World Wide Web and rapidly growing number of available digital color images, much research effort is devoted to the development of efficient content-based image retrieval systems. In this paper we propose to apply the for color image indexing. Using the proposed approach, the color histograms are being modelled as a sum of Gaussian distributions and their parameters serve as signatures, which provide for fast and efficient color image retrieval. The results of the performed experiments show that the proposed approach is robust to color image distortions introduced by lossy compression artifacts and therefore it is well suited for indexing and retrieval of Internet based collections of color images stored in lossy compression formats.

In this paper, we present a logo and trademark retrieval system for general, unconstrained, color image databases that extends the Color Edge Co-occurrence Histogram (CECH) object detection scheme. We introduce more accurate information to the CECH, by virtue of incorporating color edge detection using vector order statistics. This produces a more accurate representation of edges in images, in comparison to the simple color pixel difference classification of edges as seen in the CECH. As such, we call this the Color Edge Gradient Co-occurrence Histogram (CEGCH). We use this as the main mechanism for our retrieval scheme. Results illustrate that the proposed retrieval system retrieves logos and trademarks with good accuracy.