Shape feature encoding via Fisher Vector for efficient fall detection in depth-videos

FV can be considered as a generalization of the BoW. In other words, BoW is a particular case of the FV. The additional gradients improve the FV's performance greatly.Smaller codebooks can be used to construct the FV, which yields lower computational cost.FV performs well even with simple linear classifiers. Elderly people, who are living alone, are at great risk if a fall event occurred. Thus, automatic fall detection systems are in demand. Some of the early automatic fall detection systems such as wearable devices has a high cost and may cause inconvenience to the daily lives of the elderly people. In this paper, an improved depth-based fall detection system is presented. Our approach uses shape based fall characterization and a Support Vector Machines (SVM) classifier to classify falls from other daily actions. Shape based fall characterization is carried out with Curvature Scale Space (CSS) features and Fisher Vector (FV) encoding. FV encoding is used because it has several advantages against the Bag-of-Words (BoW) model. FV representation is robust and performs well even with simple linear classifiers. Extensive experiments on SDUFall dataset, which contains five daily activities and intentional falls from 20 subjects, show that encoding CSS features with FV encoding and a SVM classifier can achieve an up to 88.83% fall detection accuracy with a single depth camera. This classification rate is 2% more accurate than the compared approach. Moreover, an overall 64.67% accuracy is obtained for 6-class action recognition, which is about 10% more accurate than the compared approach.