Human Action Recognition Based on Tracking Features - Foundations on Natural and Artificial Computation - page 470

Human Action Recognition Based on Tracking
Features

Javier Hernandez, Antonio S. Montemayor,
Juan Jose Pantrigo, and Angel Sanchez

Departamento de Ciencias de la Computacion
Universidad Rey Juan Carlos, C/Tulipan, s/n,
28933 Mostoles, Madrid, Spain
{ javier.hernandez,antonio.sanz,juanjose.pantrigo,angel.sanchez } @urjc.es

Abstract. Visual recognition of human actions in image sequences is
an active field of research. However, most recent published methods use
complex models and heuristics of the human body as well as to classify
their actions. Our approach follows a different strategy. It is based on
simple feature extraction from descriptors obtained from a visual track-
ing system. The tracking system is able to bring some useful information
like position and size of the subject at every time step of a sequence,
and in this paper we show that, the evolution of some of these features
is enough to classify an action in most of the cases.

1 Introduction

Human action recognition aims to understand patterns of human movement
from image sequences and classify those actions into known categories. This is a
relevant problem in computer vision since it has applications in video surveillance
and monitoring human-computer interactions, augmented reality, and so on [1].
Human actions consist of spatial-temporal patterns that are generated by a
complex and time varying non-linear dynamic system. A complete description
of the system requires enumeration of all the variables, their interdependencies,
equations controlling their evolution and a set of boundary conditions to be
satisfied by the system [9]. Usually, the processing of this description needs too
many computational resources becoming intractable in real time for most of the
cases.
An standard approach for human action recognition is to extract a set of
features from each image sequence and use it to train classifiers to perform
recognition. Using those properties a system can classify or approximate a model
and use this model to classify. Approaches can be grouped depending on the
image properties such as motion-based, shape-based, gradient-based, etc. [3].
Several features have been proposed in the literature. In Wang and Suter [1]
some features extracted from human silhouettes or from their distance transform
are classified using three different methods: Gaussian mixture models, matching
based with the Hausdorf distance and continuous hidden Markov models. Zhou
et al. [4] employed time-space human silhouettes that are transformed into low