In many scenarios, a persons behavior in office environment needs to be monitored and some predefined abnormal actions or activities should be detected and recognized. In this paper, we attempted towards the solution starting from a persons pose with poselets as the basic building blocks. The existed powerful pose representation, i.e., poselets, together with deep convolutional neural networks, are exploited to implement an efficient action recognition system from still images. The system extends poselets detector to region proposal, cascaded with R∗CNN for final action detection. Unlike many published work which only emphases on action classification, our system implements multi-task learning with classification and localization of person and the corresponding actions simultaneously, To facilitate our studies, a specially designed action dataset was created. Preliminary experiments demonstrate promising results.