F.J. Madrid-Cuevas

Depth silhouettes for gesture recognition

This paper, evaluates the influence of depth information on the gesture recognition process. We propose depth silhouettes, a natural extension of the binary silhouette concept, as a mechanism to incorporate depth information for gesture recognition. Using depth silhouettes, we define extensions of three classic techniques employed previously for gesture recognition with monocular vision. These include: (a) silhouette compression using PCA and learning with HMM; (b) an exemplar-based gesture recognition using HMM; and (c) temporal templates that in this work are compressed using PCA and learned with SVM. The results obtained show that, independently of the technique employed, the use of depth silhouettes increases the success significantly. Additionally, we show how the best results are obtained through the combined use of PCA and HMM.

Automatic generation of consensus ground truth for the comparison of edge detection techniques

Two new methods are proposed to automatically generate consensus ground truth for real images: Minimean and Minimax methods. These methods and a version of the Yitzhaky and Peli method have been used to provide ground truth for the comparison of edge detection techniques. The developed experiments have revealed that the Minimean consensus method is suitable for the comparison of edge detectors because its results are equivalent to those obtained with artificial or manual ground truth.

Multi-camera people tracking using evidential filters

This work proposes a novel filtering algorithm that constitutes an extension of Bayesian particle filters to the Dempster-Shafer theory. Our proposal solves the multi-target problem by combining evidences from multiple heterogeneous and unreliable sensors. The modelling of uncertainty and absence of knowledge in our approach is specially attractive since it does not require to specify prior nor conditionals that might be difficult to obtain in complex problems. The algorithm is employed to propose a novel solution to the multi-camera people tracking problem in indoor environments. For each particle, the evidence of finding the person being tracked at the particle location is calculated by each sensor. Sensors also provide a degree of evidence about their reliability. The reliability is calculated based on the visible portion of the targets and their occlusions. Evidences collected from the camera set are fused considering their reliability to calculate the best hypothesis. The experiments conducted in several environments show the validity of the proposal.

People detection and tracking with multiple stereo cameras using particle filters

In this work, a novel approach for people detection and tracking using multiple stereo cameras is proposed. Our proposal consists in combining information from all the available cameras using three different plan-view maps. Occupancy and height maps register the volume and height of the objects that are visible in the stereo cameras, respectively. We also propose the use of a novel map, named confidence map, which registers the confidence of the information projected in each cell. The proposed confidence map is employed to fuse the information captured by each camera so that the most reliable information is kept in each cell. We then propose a particle filter algorithm for tracking people in the fused plan-view maps. The observation model employed considers height, occupancy and confidence information so that information from the most reliable camera is employed at each time instant. The experiments conducted show the validity of our proposal.

The computation of polygonal approximations for 2D contours based on a concavity tree

A proposal to improve methods to obtain polygonal approximations is proposed.The required levels of detail are achieved using the concavity tree.The local Measurement ISE/CR is used as stop condition.The proposed algorithm improves the methods tested. In this work, a new proposal to improve some methods based on the merge approach to obtain polygonal approximations in 2D contours is presented. These methods use a set of candidate dominant points (CDPs) to obtain a polygonal approximation. Then, redundant candidate dominant points of the set of CDPs are deleted, and the remaining dominant points will be the polygonal approximation of the original contour. The main drawback of most of these methods is that they use all breakpoints as CDPs and most of these breakpoints depict only the noise of the original contour.Our proposal, based on a concavity tree, obtains a more reduced and significant set of CDPs. When this proposal is used by some methods based on the merge approach (the Masood methods and the Carmona method), their computation times are reduced. The experimental results show that the new proposal is efficient and improves the tested methods.

The AVA Multi-View Dataset for Gait Recognition

In this paper, we introduce a new multi-view dataset for gait recognition. The dataset was recorded in an indoor scenario, using six convergent cameras setup to produce multi-view videos, where each video depicts a walking human. Each sequence contains at least 3 complete gait cycles. The dataset contains videos of 20 walking persons with a large variety of body size, who walk along straight and curved paths. The multi-view videos have been processed to produce foreground silhouettes. To validate our dataset, we have extended some appearance-based 2D gait recognition methods to work with 3D data, obtaining very encouraging results. The dataset, as well as camera calibration information, is freely available for research purposes.

An octree-based method for shape from inconsistent silhouettes

Shape-from-Silhouette (SfS) is the widely known problem of obtaining the 3D structure of an object from its silhouettes. Two main approaches can be employed: those based on voxel sets, which perform an exhaustive search of the working space, and those based on octrees, which perform a top-down analysis that speeds up the computation. The main problem of both approaches is the need for perfect silhouettes to obtain accurate results. Perfect background subtraction hardly ever happens in realistic scenarios, so these techniques are restricted to controlled environments where the consistency hypothesis can be assumed. Recently, some approaches (all of them based on voxel sets) have been proposed to solve the problem of inconsistency. Their main drawback is the high computational cost required to perform an exhaustive analysis of the working space. This paper proposes a novel approach to solve SfS with inconsistent silhouettes from an octree based perspective. The inconsistencies are dealt by means of the Dempster-Shafer (DS) theory and we employ a Butterworth function for adapting threshold values in each resolution level of the octree. The results obtained show that our proposal provides higher reconstruction quality than the standard octree based methods in realistic environments. When compared to voxel set approaches that manage inconsistency, our method obtains similar results with a reduction in the computing time of an order of magnitude.

Stereo Pictorial Structure for 2D articulated human pose estimation

In this paper, we consider the problem of 2D human pose estimation on stereo image pairs. In particular, we aim at estimating the location, orientation and scale of upper-body parts of people detected in stereo image pairs from realistic stereo videos that can be found in the Internet. To address this task, we propose a novel pictorial structure model to exploit the stereo information included in such stereo image pairs: the Stereo Pictorial Structure (SPS). To validate our proposed model, we contribute a new annotated dataset of stereo image pairs, the Stereo Human Pose Estimation Dataset (SHPED), obtained from YouTube stereoscopic video sequences, depicting people in challenging poses and diverse indoor and outdoor scenarios. The experimental results on SHPED indicates that SPS improves on state-of-the-art monocular models thanks to the appropriate use of the stereo information.

A new approach for multi-view gait recognition on unconstrained paths

We propose a new approach for multi-view gait recognition.The method is based on the analysis of 3D reconstructions of gait sequences.A new rotation invariant gait descriptor, based on 3D angular analysis is proposed.Experimental results demonstrate the effectiveness on unconstrained paths.The system can correctly identify about 99% of individuals on two public datasets. Direction changes cause difficulties for most of the gait recognition systems, due to appearance changes. We propose a new approach for multi-view gait recognition, which focuses on recognizing people walking on unconstrained (curved and straight) paths. To this effect, we present a new rotation invariant gait descriptor which is based on 3D angular analysis of the movement of the subject. Our method does not require the sequence to be split into gait cycles, and is able to provide a response before processing the whole sequence. A Support Vector Machine is used for classifying, and a sliding temporal window with majority vote policy is used to reinforce the classification results. The proposed approach has been experimentally validated on AVA Multi-View Dataset and Kyushu University 4D Gait Database and compared with related state-of-art work. Experimental results demonstrate the effectiveness of this approach in the problem of gait recognition on unconstrained paths.

Entropy volumes for viewpoint-independent gait recognition

Gait as biometrics has been widely used for human identification. However, direction changes cause difficulties for most of the gait-recognition systems, due to appearance changes. This study presents an efficient multi-view gait-recognition method that allows curved trajectories on completely unconstrained paths for indoor environments. Our method is based on volumetric reconstructions of humans, aligned along their way. A new gait descriptor, termed as gait entropy volume (GEnV), is also proposed. GEnV focuses on capturing 3D dynamical information of walking humans through the concept of entropy. Our approach does not require the sequence to be split into gait cycles. A GEnV-based signature is computed on the basis of the previous 3D gait volumes. Each signature is classified by a support vector machine, and a majority voting policy is used to smooth and reinforce the classifications results. The proposed approach is experimentally validated on the “AVA Multi-View Gait Dataset (AVAMVG)” and on the “Kyushu University 4D Gait Database (KY4D)”. The results show that this new approach achieves promising results in the problem of gait recognition on unconstrained paths.