This document summarizes an approach for unconstrained activity recognition in an office environment. Two PTZ cameras are used to track faces and recognize people, keeping them centered in the frame. An action dataset of 863 samples covering five actions was collected from two viewpoints. DenseTrack features including HOG, HOF, and MBH were extracted and an SVM classifier with k-fold cross validation achieved an average accuracy of 67% for action classification. Facial recognition using Fisherfaces was more reliable than other algorithms for tagging actions with people.