Ignasi Rius

Action-specific motion prior for efficient Bayesian 3D human body tracking

In this paper, we aim to reconstruct the 3D motion parameters of a human body model from the known 2D positions of a reduced set of joints in the image plane. Towards this end, an action-specific motion model is trained from a database of real motion-captured performances, and used within a particle filtering framework as a priori knowledge on human motion. First, our dynamic model guides the particles according to similar situations previously learnt. Then, the state space is constrained so only feasible human postures are accepted as valid solutions at each time step. As a result, we are able to track the 3D configuration of the full human body from several cycles of walking motion sequences using only the 2D positions of a very reduced set of joints from lateral or frontal viewpoints.

Importance of detection for video surveillance applications

Though it is the first step of a real video surveillance applica- tion, detection has received less attention than tracking in research on video surveillance. We show, however, that the majority of errors in the tracking task are due to wrong detection. We show this by experimenting with a multi object tracking algorithm based on a Bayesian framework and a particle filter. This algorithm, which we have named iTrack, is specifically designed to work in practical applications by defining a sta- tistical model of the object appearance to build a robust likelihood func- tion. Likewise, we present an extension of a background subtraction al- gorithm to deal with active cameras. This algorithm is used in the detection task to initialize the tracker by means of a prior density. By defining appropriate performance metrics, the overall system is evalu- ated to elucidate the importance of detection for video surveillance applications.

Improving tracking by handling occlusions

Keeping track of a target by successive detections may not be feasible, whereas it can be accomplished by using tracking techniques. Tracking can be addressed by means of particle filtering. We have developed a new algorithm which aims to deal with some particle-filter related problems while coping with expected difficulties. In this paper, we present a novel approach to handling complete occlusions. We focus also on the target-model update conditions, ensuring proper tracking. The proposal has been successfully tested in sequences involving multiple targets, whose dynamics are highly non-linear, moving over clutter.

Action Spaces for Efficient Bayesian Tracking of Human Motion

Bayesian tracking implemented as a particle filter is one of the most used techniques for full-body human tracking. However, given the high-dimensionality of the models to be tracked, the number of required particles to properly populate the space of solutions makes the problem computationally very expensive. To overcome this, we present an efficient scheme which makes use of an action model that guides the prediction step of the particle filter. In this manner, particles are propagated to locations in the search space with most a posteriori information. Hence, we sample from a smooth motion model only those postures which are feasible given a particular action. We show that this scheme improves the efficiency and accuracy of the overall tracking approach

Posture constraints for bayesian human motion tracking

One of the most used techniques for full-body human tracking consists of estimating the probability of the parameters of a human body model over time by means of a particle filter. However, given the high-dimensionality of the models to be tracked, the number of required particles to properly populate the space of solutions makes the problem computationally very expensive. To overcome this, we present an efficient scheme which makes use of an action-specific model of human postures to guide the prediction step of the particle filter, so only feasible human postures are considered. As a result, the prediction step of this model-based tracking approach samples from a first order motion model only those postures which are accepted by our action-specific model. In this manner, particles are propagated to locations in the search space with most a posteriori information avoiding particle wastage. We show that this scheme improves the efficiency and accuracy of the overall tracking approach

Nonlinear synchronization for automatic learning of 3D pose variability in human motion sequences

A dense matching algorithm that solves the problem of synchronizing prerecorded human motion sequences, which show different speeds and accelerations, is proposed. The approach is based on minimization of MRF energy and solves the problem by using Dynamic Programming. Additionally, an optimal sequence is automatically selected from the input dataset to be a time-scale pattern for all other sequences. The paper utilizes an action specific model which automatically learns the variability of 3D human postures observed in a set of training sequences. The model is trained using the public CMU motion capture dataset for the walking action, and a mean walking performance is automatically learnt. Additionally, statistics about the observed variability of the postures and motion direction are also computed at each time step. The synchronized motion sequences are used to learn a model of human motion for action recognition and full-body tracking purposes.

Robust particle filtering for object tracking

This paper addresses the filtering problem when no assumption about linearity or gaussianity is made on the involved density functions. This approach, widely known as particle filtering, has been explored by several previous algorithms, including Condensation. Although it represented a new paradigm and promising results have been achieved, it has several unpleasant behaviours. We highlight these misbehaviours and propose an algorithm which deals with them. A test-bed, which allows proof-testing of new approaches, has been developed. The proposal has been successfully tested using both synthetic and real sequences.

A 3d dynamic model of human actions for probabilistic image tracking

In this paper we present a method suitable to be used for human tracking as a temporal prior in a particle filtering framework such as CONDENSATION [5]. This method is for predicting feasible human postures given a reduced set of previous postures and will drastically reduce the number of particles needed to track a generic high-articulated object. Given a sequence of preceding postures, this example-driven transition model probabilistically matches the most likely postures from a database of human actions. Each action of the database is defined within a PCA-like space called UaSpace suitable to perform the probabilistic match when searching for similar sequences. So different, but feasible postures of the database become the new predicted poses.

3D action modeling and reconstruction for 2d human body tracking

In this paper we present a technique for predicting the 2D human body joints and limbs position in monocular image sequences, and reconstructing its corresponding 3D postures using information provided by a 3D action model. This method is used in a framework based on particle filtering, for the automatic tracking and reconstruction of the 3D human body postures. A set of the reconstructed postures up to time t are projected on the action space defined in this work, which is learnt from Motion Capture data, and provides us a principled way to establish similarity between body postures, natural occlusion handling, invariance to viewpoint, robustness, and is able to handle different people and different speeds while performing an action. Results on manually selected joint positions on real image sequences are shown in order to prove the correctness of this approach.

Hierarchical Eyelid and Face Tracking

Most applications on Human Computer Interaction (HCI) require to extract the movements of user faces, while avoiding high memory and time expenses. Moreover, HCI systems usually use low-cost cameras, while current face tracking techniques strongly depend on the image resolution. In this paper, we tackle the problem of eyelid tracking by using Appearance-Based Models, thus achieving accurate estimations of the movements of the eyelids, while avoiding cues, which require high-resolution faces, such as edge detectors or colour information. Consequently, we can track the fast and spontaneous movements of the eyelids, a very hard task due to the small resolution of the eye regions. Subsequently, we combine the results of eyelid tracking with the estimations of other facial features, such as the eyebrows and the lips. As a result, a hierarchical tracking framework is obtained: we demonstrate that combining two appearance-based trackers allows to get accurate estimates for the eyelid, eyebrows, lips and also the 3D head pose by using low-cost video cameras and in real-time. Therefore, our approach is shown suitable to be used for further facial-expression analysis.