Human Action Recognition Based on Tracking Features - Foundations on Natural and Artificial Computation - page 472

larger weight values can be chosen several times, a diffusion stage is applied to
avoid the loss of diversity. Finally, particle set for the next time step is predicted
using the motion model. The pseudo-code of PF is detailed in [15].
In short, Particle Filters are algorithms that handle the evolution of particles.
Particles are driven by the state model and are multiplied by their fitness values
as determined by the pdf [18]. In visual tracking problems, this pdf represents
the probability that the object is in a determined position.
2.2 Memetic Algorithms
The term memetic algorithm (MA) refers to a class of metaheuristics based on a
population of agents [14]. This method has been successfully applied to a variety
of optimization problems [17]. MA allows us to exploit all available knowledge
about the problem under study. That is what makes MA different from other
evolutionary methods. This philosophy is illustrated with the term ' meme ',
which denotes an analogy with the ' gene ' in the context of cultural evolution. The
ideas (' memes ') are propagated from brain to brain via the cultural processes in
a way similar to how the gene pool (' genes ') is propagated from body to body
via reproduction processes.
The key idea of the MA is the combination of individual improvement pro-
cedures with cooperation and competition processes in a populational context.
Algorithmically, MA maintains a population of agents during the whole opti-
mization process. Agents are basically solutions of the problem. These agents
are related to each other in a competition and cooperation context, which is
organized in different generations. Each generation is a new update of the pop-
ulation, which is performed by recombining the features of some selected agents
and replacing some agents with the new ones. The selection and replacement
procedures are both competitive processes, while the combination stage is a co-
operation process in which the selected agents generate new ones by applying
reproductive operators (combination and mutation). Finally, the MA allows the
application of local improvement procedures on some of the agents. The im-
provement procedure can be used at different stages of the optimization process,
for example: as a mutation operator, only at the end of the process, etc.
2.3 The Hybrid Method: Memetic Algorithm Particle Filter
MAPF is the result of hybridize the memetic algorithm (MA) as optimization
procedure and the particle filter (PF) as prediction procedure in two different
stages:
- PF stage is focused on the temporal evolution of a representative set of
solutions. A solution set, called SupportSet, of size N is propagated in time
and updated to obtain a new set in every time step. The aim when using PF
is tracking multiple hypotheses and using the knowledge about the system
dynamics for future prediction.