This one-day BMVA Technical Meeting was held on 23 October at the BCS in London and was chaired by Dimitrios Makris (Kingston University) and Aphrodite Galata (University of Manchester). The meeting concentrated on human motion recovery using computer vision techniques. A variety of areas involved in this arduous task were explored, such as detection, tracking and modelling. Thus, in the morning, oral presentations were mainly focused on human detection using lowlevel features, whereas in the afternoon the session went towards dimensionality reduction methods as the way to model human behaviour, which represents a significant trend in human modelling over the last few years.

The first speaker, Ben Daubney (University of Bristol), described a method for human pose estimation and tracking using sparse and noisy motion features. In his approach a model of human gait allows posture estimation in a monocular scenario using a Pictorial Structure method. Then, Jean-Christophe Nebel (Kingston University) presented a method for estimating human pose from an isolated single frame aiming at initialising articulated tracking algorithms. His approach was based on iterative fitting of a simple articulated
model over a spatio-temporal feature clustering using multiple image cues. The last morning speaker was Sam Johnson (University of Leeds) who proposed a  probabilistic algorithm for pose estimation under unconstrained natural images. This method combined HoG descriptor and colour cues in order to detect the limbs, also using a Pictorial Structure Model approach.

After lunch, Jesus Martinez del Rincon introduced a shape-skeleton tracking approach developed at the University of Zaragoza. The proposal tackles the problem of human pose recovery on video surveillance sequences and defends the advantages of 2D models such as the easy initialisation in uncalibrated environments. It consists of deterministic (Active shape
models) and stochastic (Particle filter) techniques in a unique framework with a feedback mechanism between them. The next talk was given by Marco Paladini (Queen Mary) and promoted the use of a framework for recovering 3D shape and motion from uncalibrated 2D
image measurements using an iterative factorisation approach. The proposed model is applied to the three different categories in which an object can be classified – rigid, deformable and/or articulated.

Spectral dimensionality reduction methods were the focus of the following two speakers. Fabio Cuzzolin discussed the advantages of using locally linear embedding (LLE) for representing 3D volumes with articulated structure and the soundness of this method against similar methods such as Isomap. Then, Michal Lewandowski (Kingston University) discussed how spectral dimensionality methods are used for modelling human activities. In particular, he showed a new variation of the classical LE algorithm that introduces spatio-temporal neighbour constraints during the creation of the dimensionality reduced space.

The last two talks of the day focused on 3D pose recovery. Aphrodite Galata (University of Manchester) described a tracking framework for calibrated multicamera scenarios based on learning complex activities. This learning is made over dimensionality reduced spaces that allow the decomposition of complex motion patterns into a vocabulary. Finally, John Darby
(Manchester Metropolitan University) described a tracking proposal based on a particle filter, which combines generative and discriminative approaches to address both known and unknown activities. 

In conclusion, this meeting has revealed intense and diverse activity in this important field of computer vision. Although convincing results have been demonstrated in specific scenarios, the estimation of articulated human motion remains a very challenging
problem for our community. Special attention should be paid to deal with real scenarios, multiple targets and unconstrained motion patterns. There is no doubt that fundamental progress will be achieved in the near future  and will provide exciting material for a new BMVA Technical Meeting on this same topic.

Dr Jesus Martinez del Rincon
Kingston University

Period01 Dec 2014

Media coverage


Media coverage