Behavior Recognition Using Multiple Depth Cameras Based on a Time-Variant Skeleton Vector Projection

User behavior recognition in a smart office environment is a challenging research task. Wearable sensors can be used to recognize behaviors, but such sensors could go unworn, making the recognition task unreliable. Cameras are also used to recognize behaviors, but occlusions and unstable lighting conditions reduce such methods' recognition accuracy. To address these problems, we propose a time-variant skeleton vector projection scheme using multiple infrared-based depth cameras for behavior recognition. The contribution of this paper is threefold: (1) The proposed method can extract reliable projected skeleton vector features by compensating occluded data using nonoccluded data; (2) the proposed occlusion-based weighting element generation can be employed to train support-vector-machine-based classifiers to recognize behaviors in a multiple-view environment; and (3) the proposed method achieves superior behavior recognition accuracy and involves less computational complexity compared with other state-of-the-art methods for practical testing environments.