Improving Action Recognition using Temporal Regions

Abstract

Recognizing actions in videos is an important task in computer vision area, having important applications such as the surveillance and assistance of the sick and disabled. Automatizing this task can improve the way we monitor actions since it is not necessary to have a human watching a video all the time. However, the classification of actions in a video is challenging since we have to identify temporal features that best represent each action. In this work, we propose an approach to obtain temporal features from videos by dividing the sequence of frames of a video into regions. Frames from these regions are merged in order to identify the temporal aspect that classifies actions in a video. Our approach yields better results when compared to a frame-by-frame classification.