Authors:
Rawya Al-Akam
and
Dietrich Paulus
Affiliation:
University of Koblenz-Landau, Germany
Keyword(s):
RGBD videos, Global Features, Local Features, Action Recognition.
Related
Ontology
Subjects/Areas/Topics:
Classification
;
Clustering
;
Feature Selection and Extraction
;
Pattern Recognition
;
Theory and Methods
Abstract:
This paper attempts to present human action recognition through the combination of local and global feature
descriptors values, which are extracted from RGB and Depth videos. A video sequence is represented as a
collection of spatio and spatio-temporal features. However, the challenging problems exist in both local and
global descriptors for classifying human actions. We proposed a novel combination of the two descriptor
methods, 3D trajectory and motion boundary histogram for the local feature and global Gist feature descriptor
for the global feature (3DTrMBGG). To solve the problems of the structural information among the local descriptors,
and clutter background and occlusion among the global descriptor, the combination of the local and
global features descriptor is used. In this paper, there are three novel combination steps of video descriptors.
First, combines motion and 3D trajectory shape descriptors. Second, extract the structural information using
global gist descriptor. Th
ird, combines these two descriptor steps to get the 3DTrMBGG feature vector from
spatio-temporal domains. The results of the 3DTrMBGG features are used along with the K-mean clustering
and multi-class support vector machine classifier. Our new method on several video actions improves performance
on actions even with low movement rate and outperforms the competing state-of-the-art -temporal
feature-based human action recognition methods.
(More)