Real-time action recognition using a multilayer descriptor with variable size

Abstract. Video analysis technology has become less expensive and more powerful in terms of storage resources and resolution capacity, promoting progress in a wide range of applications. Video-based human action detection has been used for several tasks in surveillance environments, such as forensic investigation, patient monitoring, medical training, accident prevention, and traffic monitoring, among others. We present a method for action identification based on adaptive training of a multilayer descriptor applied to a single classifier. Cumulative motion shapes (CMSs) are extracted according to the number of frames present in the video. Each CMS is employed as a self-sufficient layer in the training stage but belongs to the same descriptor. A robust classification is achieved through individual responses of classifiers for each layer, and the dominant result is used as a final outcome. Experiments are conducted on five public datasets (Weizmann, KTH, MuHAVi, IXMAS, and URADL) to demonstrate the effectiveness of the method in terms of accuracy in real time.