Abstract—Hidden Markov model is well-known for its application in temporal pattern recognition. Its disadvantages are its computational expensive and very prone to numerically underflow. The focus of this paper is on the forward algorithm which computes the probability of a particular output sequence with respect to the HMM parameters. To make it feasible to implement on an embedded system, we propose a modified forward algorithm that makes use of integer only representation and operations. The outcome of these modifications is integrated into a motion type classification system used for elderly monitoring; 7 motion types with 2×2 transition and 8 emission probabilities, on a low cost embedded system based on a 32-bit ARM Cortex-M0+. The system is capable to perform with comparable classification correctness to the ordinary and the scaling coefficients algorithms. It outperforms the ordinary ones by taking 10 percent of time, 91 percent of code size and 54 percent of memory. It is capable of forcing the processor to sleep the longest with only 3.1ms execution time per second (8.7 and 3.1 percent of the ordinary and the scaling coefficients algorithms). This makes it more suitable for implementation on an embedded system.

 

Index Terms—hidden markov model, embedded system, motion type classification, forward algorithm, elderly monitoring system

 

Cite: W. Kurdthongmee, "A Modified HMM Forward Algorithm for an Embedded Motion Type Classification," International Journal of Signal Processing Systems, Vol. 2, No. 2, pp. 83-90, December 2014. doi: 10.12720/ijsps.2.2.83-90