Carlos Orrite

Randomized trees for human pose detection

This paper addresses human pose recognition from video sequences by formulating it as a classification problem. Unlike much previous work we do not make any assumptions on the availability of clean segmentation. The first step of this work consists in a novel method of aligning the training images using 3D Mocap data. Next we define classes by discretizing a 2D manifold whose two dimensions are camera viewpoint and actions. Our main contribution is a pose detection algorithm based on random forests. A bottom-up approach is followed to build a decision tree by recursively clustering and merging the classes at each level. For each node of the decision tree we build a list of potentially discriminative features using the alignment of training images; in this paper we consider Histograms of Orientated Gradient (HOG). We finally grow an ensemble of trees by randomly sampling one of the selected HOG blocks at each node. Our proposed approach gives promising results with both fixed and moving cameras.

Shape matching of partially occluded curves invariant under projective transformation

This paper describes a method to identify partially occluded shapes which are randomly oriented in 3D space. The goal is to match the object contour present in an image with an object in a database. The approach followed is the alignment method which has been described in detail in the literature. Using this approach the recognition process is divided into two stages: first, the transformation between the viewed object and the model object is determined, and second, the model that best matches the viewed object is found. In the first stage, invariant points under projective transformation (based on bitangency) are used, which drastically reduced the selection space for alignment. Next, the curves are compared after the transformation matrix is estimated between the image and the model in order to determine the pose of the curve that undergoes the perspective projection. The evaluation process is performed using a novel estimation of the Hausdorff distance (HD), called the continuity HD. It evaluates partially occluded curves in the image in relation to the complete contour in the database. The experimental results showed that the present algorithm can cope with noisy figures, projective transformations, and complex occlusions. 2003 Elsevier Inc. All rights reserved.

HOG-Based Decision Tree for Facial Expression Classification

We address the problem of human emotion identification from still pictures taken in semi-controlled environments. Histogram of Oriented Gradient (HOG) descriptors are considered to describe the local appearance and shape of the face. First, we propose a Bayesian formulation to compute class specific edge distribution and log-likelihood maps over the entire aligned training set. A hierarchical decision tree is then built using a bottom-up strategy by recursively clustering and merging the classes at each level. For each branch of the tree we build a list of potentially discriminative HOG features using the log-likelihood maps to favor locations that we expect to be more discriminative. Finally, a Support Vector Machine (SVM) is considered for the decision process in each branch. The evaluation of the present method has been carried out on the Cohn-Kanade AU-Coded Facial Expression Database, recognizing different emotional states from single picture of people not present in the training set.

Classifier Ensemble Generation for the Majority Vote Rule

This paper addresses the problem of classifier ensemble generation. The goal is to obtain an ensemble to achieve maximum recognition gains with the lowest number of classifiers. The final decision is taken following a majority vote rule. If the classifiers make independent errors, the majority vote outperforms the best classifier. Therefore, the ensemble should be formed by classifiers exhibiting individual accuracy and diversity. To account for the quality of the ensemble, this work uses a sigmoid function to measure the behavior of the ensemble in relation to the majority vote rule, over a test labelled data set.

Rao-Blackwellised particle filter for colour-based tracking

Colour-based particle filters have been used exhaustively in the literature, given rise to multiple applications. However, tracking coloured objects through time has an important drawback, since the way in which the camera perceives the colour of the object can change. Simple updates are often used to address this problem, which imply a risk of distorting the model and losing the target. In this paper, a joint image characteristic-space tracking is proposed, which updates the model simultaneously to the object location. In order to avoid the curse of dimensionality, a Rao-Blackwellised particle filter has been used. Using this technique, the hypotheses are evaluated depending on the difference between the model and the current target appearance during the updating stage. Convincing results have been obtained in sequences under both sudden and gradual illumination condition changes.

Tracking human body parts using particle filters constrained by human biomechanics

In this paper, a novel framework for visual tracking of human body parts is introduced. The presented approach demonstrates the feasibility of recovering human poses with data from a single uncalibrated camera using a limb tracking system based on a 2D articulated model. It is constrained only by biomechanical knowledge about human bipedal motion, instead on relying on constraints linked to a specific activity or camera view. These characteristics make our approach suitable for real visual surveillance applications. Experiments on HumanEva I & II datasets demonstrate the effectiveness of our method on tracking human lower body parts. Moreover, a detail comparison with current tracking methods is presented.

View-invariant human feature extraction for video-surveillance applications

We present a view-invariant human feature extractor (shape+pose) for pedestrian monitoring in man-made environments. Our approach can be divided into 2 steps: firstly, a series of view-based models is built by discretizing the viewpoint with respect to the camera into several training views. During the online stage, the Homography that relates the image points to the closest and most adequate training plane is calculated using the dominant 3D directions. The input image is then warped to this training view and processed using the corresponding view-based model. After model fitting, the inverse transformation is performed on the resulting human features obtaining a segmented silhouette and a 2D pose estimation in the original input image. Experimental results demonstrate our system performs well, independently of the direction of motion, when it is applied to monocular sequences with high perspective effect.

Monocular 3-D Gait Tracking in Surveillance Scenes

Gait recognition can potentially provide a noninvasive and effective biometric authentication from a distance. However, the performance of gait recognition systems will suffer in real surveillance scenarios with multiple interacting individuals and where the camera is usually placed at a significant angle and distance from the floor. We present a methodology for view-invariant monocular 3-D human pose tracking in man-made environments in which we assume that observed people move on a known ground plane. First, we model 3-D body poses and camera viewpoints with a low dimensional manifold and learn a generative model of the silhouette from this manifold to a reduced set of training views. During the online stage, 3-D body poses are tracked using recursive Bayesian sampling conducted jointly over the scenes ground plane and the pose-viewpoint manifold. For each sample, the homography that relates the corresponding training plane to the image points is calculated using the dominant 3-D directions of the scene, the sampled location on the ground plane and the sampled camera view. Each regressed silhouette shape is projected using this homographic transformation and is matched in the image to estimate its likelihood. Our framework is able to track 3-D human walking poses in a 3-D environment exploring only a 4-D state space with success. In our experimental evaluation, we demonstrate the significant improvements of the homographic alignment over a commonly used similarity transformation and provide quantitative pose tracking results for the monocular sequences with a high perspective effect from the CAVIAR dataset.

Exploiting projective geometry for view-invariant monocular human motion analysis in man-made environments

Example-based approaches have been very successful for human motion analysis but their accuracy strongly depends on the similarity of the viewpoint in testing and training images. In practice, roof-top cameras are widely used for video surveillance and are usually placed at a significant angle from the floor, which is different from typical training viewpoints. We present a methodology for view-invariant monocular human motion analysis in man-made environments in which we exploit some properties of projective geometry and the presence of numerous easy-to-detect straight lines. We also assume that observed people move on a known ground plane. First, we model body poses and silhouettes using a reduced set of training views. Then, during the online stage, the homography that relates the selected training plane to the input image points is calculated using the dominant 3D directions of the scene, the location on the ground plane and the camera view in both training and testing images. This homographic transformation is used to compensate for the changes in silhouette due to the novel viewpoint. In our experiments, we show that it can be employed in a bottom-up manner to align the input image to the training plane and process it with the corresponding view-based silhouette model, or top-down to project a candidate silhouette and match it in the image. We present qualitative and quantitative results on the CAVIAR dataset using both bottom-up and top-down types of framework and demonstrate the significant improvements of the proposed homographic alignment over a commonly used similarity transform.

Mean field approach for tracking similar objects

In this paper, we consider the problem of tracking similar objects. We show how a mean field approach can be used to deal with interacting targets and we compare it with Markov Chain Monte Carlo (MCMC). Two mean field implementations are presented. The first one is more general and uses particle filtering. We discuss some simplifications of the base algorithm that reduce the computation time. The second one is based on suitable Gaussian approximations of probability densities that lead to a set of self-consistent equations for the means and covariances. These equations give the Kalman solution if there is no interaction. Experiments have been performed on two kinds of sequences. The first kind is composed of a single long sequence of twenty roaming ants and was previously analysed using MCMC. In this case, our mean field algorithms obtain substantially better results. The second kind corresponds to selected sequences of a football match in which the interaction avoids tracker coalescence in situations where independent trackers fail.