Human Action Recognition Based on Tracking Features - Foundations on Natural and Artificial Computation - page 473

- In the MA stage, solutions are combined and improved to obtain new better
ones using the strategies of the Memetic Algorithm metaheuristic. This stage
is a refinement step embedded in the PF framework and is based on a set
of procedures extracted from MA general schema: selection, combination,
mutation and improvement.
MAPF addresses the search to regions of the solution space in which it is highly
probable to find new better solutions. MA stage performs a rational search be-
yond the simple stochastic procedure used by PF. On the other hand PF stages
increase the performance of general optimization algorithms in dynamic prob-
lems by improving the quality of the diverse initial solution set. MA and PF are
related in such a way that when MA improves its results, PF performance also
improves, and vice versa.
MAPF can handle different size state-space representation in each stage ac-
cording to the number of estimated features. In our case we track a single object
and estimate its bounding box [ x, y, Lx, Ly ] represented by position and size.
It is possible to estimate an approximate position using the PF stage, so the
PF stage determines the position [ x, y ] resulting in a two dimensional problem.
Memetic Algorithm refines this estimation considering the size of the target in
this case the associate state-space of the MA is [ x, y, Lx, Ly ]. This is very use-
ful to reduce the workload of the particle filter when using high dimensional
state-spaces.
Fig. 1. Algorithmic details of the MAPF as the core of the tracking module
2.4 Structure of MAPF Output
The MAPF manages a population of solutions where each solution is stored in a
particle. Each solution contains the set of variables which describe the system-
state and its weight. As a result, the proposed state-space model for object
tracking is 4 dimensional space. The structure which stores a solution is a state