Silhouette based human fall detection using multimodal classifiers for content based video retrieval systems

摘要

Automatic human fall detections are proving the need of time in emergency situations for elderly persons falling on the floor injuring them, sometimes with bone fracture or more severely at times being alone while performing their daily activities. Recent advancement in image processing and therein activity identification is seeing rising trend of research. Present paper aims at putting forward a fall detection system which uses human silhouettes, as processed from depth cue based on camera footages, to extract curvature scale space (CSS) features. Human actions thus finally rendered into CSS are classified with the help of standard machine learning classifier techniques such as support vector machine (SVM) and extreme learning machine (ELM). Moreover, the paper distinctively puts forward the benefit of augmented ELM classifier with help of sparse representation for image frame classification (SRC) technique. The system had been tested with standard dataset as established in literature for human action classification. The results presented in form of confusion matrix comprising of detecting semantic activities like walking, idle sitting, standing and falling demonstrate that the developed system has an edge in terms of higher accuracy compared to similar state of the art methods as reported in literature.