Unsupervised learning spatio-temporal features for human activity recognition from RGB-D video data

Being able to recognize human activities is essential for several applications, including social robotics. The recently developed commodity depth sensors open up newpossibilities of dealingwith this problem. Existing techniques extract hand-tuned features, such as HOG3D or STIP, from video data. They are not adapting easily to new modalities. In addition, as the depth video data is lowquality due to the noise, we face a problem: does the depth video data provide extra information for activity recognition? To address this issue, we propose to use an unsupervised learning approach generally adapted to RGB and depth video data. we further employ the multi kernel learning (MKL) classifier to take into account the combinations of different modalities. We show that the low-quality depth video is discriminative for activity recognition. We also demonstrate that our approach achieves superior performance to the state-of-the-art approaches on two challenging RGB-D activity recognition datasets. © Springer International Publishing 2013