Human Action Recognition Based on Tracking Features - Foundations on Natural and Artificial Computation - page 474

vector s t =[ x t ,y t ,Lx t ,Ly t ], where ( x t ,y t ) represents the position of the object
o geometrical center in the global image frame at time t ,and( Lx t ,Ly t )isthe
height and width of a bounding box fitting the object o in the frame at time t .

3 Feature Extraction

An standard approach of a human action recognition system is to extract a set
of features from each image sequence and use it to train a classifier that will
eventually perform the recognition. Therefore, a key point in this approach is to
select a robust feature set which allows to accurately discriminate the considered
actions.
Feature selection is based on expert human knowledge. This model has a
significant interest because it relies directly on an expression of facts stated in
natural language. The model can be validated and allows to check the consistency
of the set of rules. Furthermore, the model can be progressively completed by
asking new questions in natural language.
Here is an example illustrating some knowledge used to select features:

- When someone is jumping in place there is no displacement
- The bend down action implies change in the height of the bounding box
- The jack action produces changes in the bounding box height and width
- While a subject walks the center of mass varies in a known direction

Considering the kind of rules stated above the idea is to find some features that
can be extracted from the subject bounding box evolution through the time.
The human knowledge will be then translated to rules involving these features.
Variances, covariances, means, maximums and minimums of the variables that
define bounding boxes are used to represent concepts as can be seen in Table 1.
In order to test that those measures are discriminant enough a classification tree
using fixed thresholds was built. The classification tree was tested against the
database found in [11] and the classification was successful. There are measurable
relations between these variables and the action being performed. An example
ofapersonwalkingisdescribedinFigure2where Lx shows a large variance
while Ly remains almost constant.

Table 1. Some used variables and their meanings related to human knowledge state-
ments about actions

Variable Meaning

var(x) displacement

var(y) jump

var(Lx) amplitude of the movement

var(Ly) change of height

cov(Lx, Ly) arm-leg coordination

cov(y, Lx) arm-leg coordination