2d human motion estimation modeling for classification
Loading...
Date
2016-04-01
Authors
Chan; Choon Kit
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Human motion estimation is an approach to predict motion activities from
static body postures; widely explored from gait motion, silhouette-based,
biomechanical-based or image-based analyses for motion capture, recognition and
vision surveillance purposes. Human motion is often captured via Marker-Based
(MB) and Marker-Less (ML) system by using single or multiple cameras. These
motions are commonly analyzed in 3-Dimensional (3D) or 2-Dimensional (2D)
positioning involving location and orientations of body joints. Nevertheless, owing to
the complexity of high dimensionality motion data, this study has focused on the 2D
human motion. Existing developed 2D stick figures could hardly point the exact
body joint location. Besides, no researchers have considered the tolerance adjustment for human motion estimation. Therefore, the main goal of this study is to develop a 2D stick estimation model with error tolerance to represent human motions for classification analysis. The 2D stick estimation model is developed from three
fundamental body segments: Backbone (BB), Upper Body (UB) and Lower Body
(LB). Considering the capability of body segments’ stretches while performing
different activities, tolerance model is derived from the average deviations of
polynomial fitting coefficients evaluated at sequential time steps. Integrating the
precedent time-step coordinates with the tolerance model iteratively yield the
estimated body joint coordinates at subsequent time step. The developed model is
tested on (i) MB basic motions: walking, running, jumping and MB sports motions:
punching, sword playing and taichi from CMU database and (ii) ML basic motions:
walking, running, jumping from YouTube and ML sports motions: Yoga motion of
child’s, leg lock and camel pose from experimental captures. Data transformation is
initiated to snapshot the video data into still images followed by image
transformations into coordinate data. Data elimination cum regression imputation is
carried out to treat missing data found from occlusion and hidden body segments.
The motion estimation model for tolerance consideration is performed on three 2D
motion estimation techniques: IVE-SAT, AVE-SAT and AVE-TAT. The 2D stick
estimation model is judged on matching analysis and classification accuracies using
Lazy classifiers. Findings show that the developed 2D stick estimation model by
AVE-TAT resulted in best matching accuracy up to 66.67% and classification
accuracies above 90% for all motion categories. The developed model has the
advantage over its ability to estimate human motions specifically with error tolerance
adjustment resembling the body segment stretches throughout the entire activity. The
study outcomes successfully imply that the proposed 2D stick estimation model with
AVE-TAT is a feasible approach in distinguishing characteristics of different human
motions for classifications.