Focus-of-Attention for Human Activity Recognition from UAVs

This paper presents a system to extract metadata about human activities from full-motion video recorded from a UAV. The pipeline consists of these components: tracking, motion features, representation of the tracks in terms of their motion features, and classification of each track as one of the human activities of interest. We consider these activities: walk, run, throw, dig, wave. Our contribution is that we show how a robust system can be constructed for human activity recognition from UAVs, and that focus-of-attention is needed. We find that tracking and human detection are essential for robust human activity recognition from UAVs. Without tracking, the human activity recognition deteriorates. The combination of tracking and human detection is needed to focus the attention on the relevant tracks. The best performing system includes tracking, human detection and a per-track analysis of the five human activities. This system achieves an average accuracy of 93%. A graphical user interface is proposed to aid the operator or analyst during the task of retrieving the relevant parts of video that contain particular human activities. Our demo is available on YouTube.