Maximum Entropy Markov Model for Human Activity Recognition Using Depth Camera

Activity recognition is an essential factor in the determination of daily routine of a human being. There exist numerous Human Activity Recognition (HAR) systems; however, an HAR with a practical accuracy is still in search and a challenge at large. The classifiers utilized in existing systems offer degrading recognition rates in various key environments such as depth camera environment among others. To address the limitation of degrading recognition rates, in this paper, we propose Maximum Entropy Markov Model (MEMM) that solves the degrading recognition rate problem. In MEMM, we model the states of activity recognition as the states of the model itself i.e. as the states of MEMM, and hence consider the observations of video-sensor as the observations of MEMM. Further, we use a modified version of Viterbi, a machine learning algorithm, to generate the most likely probable state sequence based on these observations. Then, from such a state sequence, we use MEMM to predict the activity state. We evaluate the performance MEMM against depth dataset having eleven different types of activities in a large-scale experimentation process. The results show that MEMM outperforms existing well-known methods by achieving a weighted average recognition rate 96.3% across the naturalistic dataset collected using depth camera.