|dc.description.abstract||In this dissertation, we address the problem of recognizing human-action
from videos. The recognition aims at recovering action information from the
image sequences using different features such as variations of the human shape.
Approaches based on such features often use sequence-alignment methods. We
propose two novel methods for human-action recognition. We also propose an
elliptical-shaped band for the Dynamic Time Warping (DTW) that provides a good
compromise between alignment accuracy and computational speed.
First, we study the applicability of the pairwise shape-similarity measurements
for human-action recognition. Since action can be seen as a sequence of shapes
of silhouette poses, there exists similarities between actions from the same class.
Based on this observation, we propose a new method for classifying human actions.
Given two sequences of silhouettes representing an action, we measure their similarity by means of a robust sequence-alignment method. The motion cue is
implicitly represented by the implicit variations of the human’s shape over time while
an action is performed. We adopt the Longest Common Sub-Sequence (LCSS), a
dynamic-programming approach that calculates the minimum cost of aligning the
Next, we use information from inter-pose shape variations as provided by shape
descriptors for recognizing human actions. Here, in contrast to the previous
method, where an action is not modeled by itself, we present a method that converts
an action into a sequence based on the variations of a human’s shape over time. We
construct the sequence using the Inner-Distance Shape-Context as a measurement
of variations between shapes. Experimental results compare our method favorably
with related methods.
Finally, we develop a new global band for the Dynamic Time Warping algorithm.
In contrast with standard rectangular-shaped bands, we propose an elliptical-shaped
band that provides flexibility and a good compromise between alignment accuracy
and computational speed. The shape of the ellipse is implicitly represented by the
length of the time series. The idea of our elliptical band is to speed up DTW
and enforce a global constraint on the warping path by using a window size that
tolerates a significant amount of noise in the aligned time series.||en_US