Recently, human action recognition in videos attracts increasing research interests in computer vision. As a result, visual features are becoming crucial for human action recognition in videos. In this paper, we propose Space-Time Binarized Normed Gradients (ST-BING) as a new feature for action recognition and localization. This feature comprises both the static information and the non-static information of an action performer. We use our ST-BING feature extracted from human action bounding

more »

... s to learn an action recognition and localization model in a cascaded Support Vector Machine (SVM) framework. The binarized version of our feature is easy to extract with only a few atomic operations. Using our feature and just trained by a simple linear SVM, we attain better than state-of-the-art action recognition performance on a challenging dataset. At the same time, our method produces good action localization results.