On the use of local motion information for human action recognition via feature selection

Automated recognition of human activities has received considerable attention within the computer vision community. This is mainly due to the plethora of applications where human activity recognition can be deployed such as smart automated surveillance and human computer interaction. In this research study, a motion descriptor is employed for the extraction of features across consecutive frames for the classification of human activities. A histogram of features is constructed from the image taking into account the solely local properties embedded within the motion map. Feature selection based on the proximity of instances belonging to the same class is applied to derive the most discriminative features. Experimental results carried out on the Weizmann dataset confirmed the potency for the proposed method to better distinguish between different activity classes such as running, walking, waving and jumping. The dataset is made of 19 basic actions for 9 different subjects.