Catálogo de publicaciones - libros

Pan-tilt cameras are widely used in surveillance networks. These cameras are often equipped with telephoto lenses to capture objects at a distance. Such a camera makes full-metric calibration more difficult since the projection with a telephoto lens is close to orthographic. This paper discusses the problems caused by pan-tilt cameras with long focal length and presents a method to improve the calibration accuracy. Experiments show that our method reduces the re-projection errors by an order of magnitude compared to popular homography-based approaches.

The projective geometric properties of two principal-axes aligned (PAA) conics in a model plane are investigated in this paper by utilized the generalized eigenvalue decomposition (GED). We demonstrate that one constraint on the image of the absolute conic (IAC) can be obtained from a single image of two PAA conics even if their parameters are unknown. And if the eccentricity of one of the two conics is given, two constraints on the IAC can be obtained. An important merit of the algorithm using PAA is that it can be employed to avoid the ambiguities when estimating extrinsic parameters in the calibration algorithms using concentric circles. We evaluate the characteristics and robustness of the proposed algorithm in experiments with synthetic and real data.

In this paper, we propose a practical intrusion detection system using uncalibrated multiple cameras. Our algorithm combines the contour based multi-planar visual hull method and a projective reconstruction method. To set up the detection system, no advance knowledge or calibration is necessary. A user can specify points in the scene directly with a simple colored marker, and the system automatically generates a restricted area as the convex hull of all specified points. To detect an intrusion, the system computes intersections of an object and each sensitive plane, which is the boundary of the restricted area, by projecting an object silhouette from each image to the sensitive plane using 2D homography. When an object exceeds one sensitive plane, the projected silhouettes from all cameras must have some common regions. Therefore, the system can detect intrusion by any object with an arbitrary shape without reconstruction of the 3D shape of the object.

We propose a fast algorithm to estimate background models using Parzen density estimation in non-stationary scenes. Each pixel has a probability density which approximates pixel values observed in a video sequence. It is important to estimate a probability density function fast and accurately. In our approach, the probability density function is partially updated within the range of the window function based on the observed pixel value. The model adapts quickly to changes in the scene and foreground objects can be robustly detected. In addition, applying our approach to cast-shadow modeling, we can detect moving cast shadows. Several experiments show the effectiveness of our approach.

This paper presents a novel color-based method to detect road signs directly from videos. A road sign usually has specific colors and high contrast to its background. Traditional color-based approaches need to train different color detectors for detecting road signs if their colors are different. This paper presents a novel color model derived from Karhunen-Loeve(KL) transform to detect road sign color pixels from the background. The proposed color transform model is invariant to different perspective effects and occlusions. Furthermore, only one color model is needed to detect various road signs. After transformation into the proposed color space, a RBF (Radial Basis Function) network is trained for finding all possible road sign candidates. Then, a verification process is applied to these candidates according to their edge maps. Due to the filtering effect and discriminative ability of the proposed color model, different road signs can be very efficiently detected from videos. Experiment results have proved that the proposed method is robust, accurate, and powerful in road sign detection.

In this paper, we propose a Semi-Supervised Multiple-Instance Learning (SSMIL) algorithm, and apply it to Localized Content-Based Image Retrieval(LCBIR), where the goal is to rank all the images in the database, according to the object that users want to retrieve. SSMIL treats LCBIR as a Semi-Supervised Problem and utilize the unlabeled pictures to help improve the retrieval performance. The comparison result of SSMIL with several state-of-art algorithms is promising.

We develop an object detection method combining top-down recognition with bottom-up image segmentation. There are two main steps in this method: a hypothesis generation step and a verification step. In the top-down hypothesis generation step, we design an improved Shape Context feature, which is more robust to object deformation and background clutter. The improved Shape Context is used to generate a set of hypotheses of object locations and figure-ground masks, which have high recall and low precision rate. In the verification step, we first compute a set of feasible segmentations that are consistent with top-down object hypotheses, then we propose a (FPP) procedure to prune out false positives. We exploit the fact that false positive regions typically do not align with any feasible image segmentation. Experiments show that this simple framework is capable of achieving both high recall and high precision with only a few positive training examples and that this method can be generalized to many object classes.

Text appearing in video provides semantic knowledge and significant information for video indexing and retrieval system. This paper proposes an effective method for text detection in video based on the similarity in stroke width of text (which is defined as the distance between two edges of a stroke). From the observation that text regions can be characterized by a dominant fixed stroke width, edge detection with local adaptive thresholds is first devised to keep text- while reducing background-regions. Second, morphological dilation operator with adaptive structuring element size determined by stroke width value is exploited to roughly localize text regions. Finally, to reduce false alarm and refine text location, a new multi-frame refinement method is applied. Experimental results show that the proposed method is not only robust to different levels of background complexity, but also effective to different fonts (size, color) and languages of text.

In this paper, a multiview pedestrian detection method based on Vector Boosting algorithm is presented. The Extended Histograms of Oriented Gradients (EHOG) features are formed via dominant orientations in which gradient orientations are quantified into several angle scales that divide gradient orientation space into a number of dominant orientations. Blocks of combined rectangles with their dominant orientations constitute the feature pool. The Vector Boosting algorithm is used to learn a tree-structure detector for multiview pedestrian detection based on EHOG features. Further a detector pyramid framework over several pedestrian scales is proposed for better performance. Experimental results are reported to show its high performance.

We propose a novel learning strategy called Global-Local Motion Pattern Classification (GLMPC) to localize pedestrian-like motion patterns in videos. Instead of modeling such patterns as a single class that alone can lead to high intra-class variability, three meaningful partitions are considered - left, right and frontal motion. An AdaBoost classifier based on the most discriminative eigenflow weak classifiers is learnt for each of these subsets separately. Furthermore, a linear three-class SVM classifier is trained to estimate the global motion direction. To detect pedestrians in a given image sequence, the candidate optical flow sub-windows are tested by estimating the global motion direction followed by feeding to the matched AdaBoost classifier. The comparison with two baseline algorithms including the degenerate case of a single motion class shows an improvement of 37% in false positive rate.