Human Action Recognition Based on Tracking Features - Foundations on Natural and Artificial Computation - page 471

dimensional multivariate time series with Tensor Subspace Analysis in order
to classify with a Gaussian process, Support Vector Machines (SVM), etc. Ali
et al. [9] used a combination of chaotic invariants as feature vector that are
classified using K-nearest neighbours. A keypose based approach where each
action is modelled as series of synthetic poses and constraints on transitions
between them is used in [5]. In [7] each action is represented as a unique curve
in a 3D invariance-space and matched against candidate action volumes in a
probabilistic framework. Finally optical flow volumetric features have also been
studied in [8].
We propose the use of a visual tracking system in order to get the height and
width as well as the position of a bounding box containing the subject in every
time step of a sequence. The evolution in time of these bounding box properties
are used as the input of a SVM classifier. The visual tracking task is performed
by a variant of a particle filter [12] that includes a memetic algorithm [14] as a
refinement stage. We called this method as Memetic Algorithm Particle Filter
(MAPF). For a detailed description of MAPF please refer to [16]. The rest of the
paper is organized as follows. Section 2 presents the MAPF as well as the PF
and MA algorithms. Section 3 describes the feature extraction module. Section 4
presents the experimental results. Finally, section 5 summarizes the conclusions
and future lines of research.
2 The MAPF Algorithm
2.1 Particle Filters for Visual Tracking
Particle Filters approximate theoretical distributions in the state-space by sim-
ulated random measures called particles [18]. The state-space model consists of
two processes: (i) observation process in which observation vector is obtained,
and (ii) a transition process. The goal is to estimate the new system state at
each time step.
In the framework of Sequential Bayesian Modelling, the posterior pdf is es-
timated in two stages (i) evaluation in which the posterior is computed using
the observation vector and (ii) propagation where the posterior is propagated at
time step t .
A predefined system model is used to update the particle set. The problem
lies in a state modelling where the dynamics equation describes the evolution of
the object and the measurement equation links the observation with the state
vector.
The aim is the recursive estimation of the posterior pdf , that constitutes a
complete solution to the sequential estimation problem. This pdf is represented
by a set of weighted particles. Each particle stores a system state at a given time
and a quality measure, proportional to the probability of the state in represents.
The PF algorithm starts by initializing a population vector using a known pdf .
The measurement vector at a given time step is obtained from the system, and
particle weights are computed using a fitness function. The weights are normal-
ized and a new particle set is selected. Taking into account that particles with