Automatic Classification of Human Body Postures Based on the Truncated SVD

Abstract

Abstract-In this experimental study, we propose the use of Singular Value Decomposition (SVD) coefficients as features to automatically classify human body postures. The classification process uses images extracted from a fixed camera video. A background subtraction technique is applied for human body segmentation. A truncated SVD is performed by selecting significant magnitude coefficients. And the height-width ratio of the human body is also included in the set of features. The classification is then performed using an Artificial Neural Network (ANN). Four body postures are considered in our experiments, namely: standing, bending, sitting, and lying. Evaluation results show that the proposed method achieved 90.46% classification accuracy. Truncated SVD coefficients and height-width ratio as body posture features are thus appropriate descriptors to achieve high classification accuracy. Also, the proposed method yields the best classification accuracy compared to well-known classification methods. Index Terms-Human body postures, classification, SVD coefficients, neural network. I. INTRODUCTIONHuman behavior understanding remains an interesting study area in intelligent video monitoring because it is closely related to computer vision applications and automatic surveillance. It is a very challenging and complex task due to some difficulties such as variations in people's appearance, skin color, illumination conditions, and the amount of data generated. To reduce the complexity, many previous works have focused on some specific topics such as face recognition [1], body tracking [2], and hand gesture recognition [3] from a single image or multiple images.Behavior understanding includes mainly motion pattern identification or body posture analysis and classification. Many approaches have been proposed for human posture classification. Poppe et al. [4] use a degree of freedom to estimate the body posture for a presenter in meeting environments. Hooi et al. propose a computer vision-based approach to automatically detect human body parts and estimate the human body postures from a monocular video sequence [5].However, most of the stated methods use skin color to locate hands and head. The person is also supposed to face the camera which can be considered as a limitation.Capo et al. present a computer vision algorithm to automatically model the human body skeleton using Bayesian networks classifier [6]. In [7] the body is decomposed into its Manuscript received August 25, 2013; revised October 29, 2013. The authors are with LCPTS laboratory, University of Sciences and Technology Houari Boumé dienne, Algeria (e-mail: nzerrouki@usthb.dz, ahouacine@usthb.dz). main parts, it constructs a silhouette based body model to determine the location of the body parts [8]. But recognition of human parts is not an easy task as it requires processing based on only visible parts of the body. Lutz Goldmann et al.[9] use appearance-based features for posture description such as color, shape, texture, motion, and size. Fast Fourier Tran...

Research on the application of body posture action feature extraction and recognition comparison

“…Through the relationship between the skeletal systems shown above, where C1_x and C1_y represent the x and y coordinates of the point C1 calculated by the two points P1 and P2, then the distance between the bones can be calculated by formula (4). And the calculation result is given to the display value by formula (5), where Distance is the distance between bones calculated by formula (4), and Th is the defined threshold variable.…”

“…In recent years, there has been continuous research on action recognition methods. Through supervised machine learning methods, action features are marked to predict unknown data for feature recognition [4], and unsupervised machine learning methods can also be used to apply them into unlabelled data [5]. Gesture action features are determined by correlation calculation and defined as feature templates, which are based on various feature templates extracted from images and videos, and body gesture action recognition is performed by matching these templates with the template database [6].…”

Human activity recognition using 2D skeleton data and supervised machine learning

“…The second type is machine learning approaches which can be subdivided into supervised and unsupervised methods. Supervised approaches use labelled data for predicting the labels of unknown data [8]. Unsupervised techniques are used for unlabelled data [9].…”