Human fall detection using CHLAC features with skeletal image sequences

Abstract

This paper describes a human fall detection approach using CHLAC features with skeletal image sequences obtained from Kinect on a mobile robot. In the proposed approach, the spatio-temporal local geometric features of moving-image sequences are extracted with CHLAC (the cubic higher-order local auto-correlation) method firstly. Then the 251-dimensional CHLAC feature vectors are projected to a low-dimensional eigenspace by principal components analysis method. The skeletal image sequences of 3 kinds of motions (walking and lying down as normal motion, fall as abnormal motion) are learned in the low-dimensional eigenspace and the basis of the eigenspace can be obtained off-line. The CHLAC feature vectors of skeleton image sequences obtained in real time are projected to the eigenspace obtained off-line. The degree of similarity to the 3 kinds of motions is evaluated and the motion is classified by k-Nearest Neighbor algorithm. The performance of the classification is evaluated with F-measure.