This document describes a system called CogWatch that uses parallel hidden Markov model (HMM) detectors and generative/discriminative models to recognize human activities for rehabilitation of stroke patients. CogWatch instruments objects like a tea set to recognize subgoals of making a cup of tea. It uses sensors on objects and a Kinect camera. Parallel HMM detectors address overlapped subgoals. The HMMs model inter- and intra-user variability in speed and sequences. Output probabilities are modeled using deep neural networks and Gaussian mixture models. Real-time decoding is achieved using a partial traceback algorithm. Experimental results show error reductions of up to 79% for recognizing subgoals using the DNN-HMM system compared to the GMM-HMM