Study And Implementation Of Human Action Recognition Based On Semi-Connected HMM

The HMM based human action recognition(HAR) approach has been broadly adopted by HAR community.Usually,the classical HMM is adopted which is constructed in the form of stochastic state transition with all states fully interconnected, namely Full-Connected HMM(FCHMM).However,existing works do not pay attention to the relationship between the layout of the model and the property of human action.To deal with the problem,we observe the property of human actions:each action can be essentially described by three key postures located around the initial,middle and terminal position of the action period;the other postures only act as transitional function between key postures.This property can be used as a plausible cue to determine the number of HMM states and their interconnections for HAR task.In this paper,we assume a weighted left-to-right three-state HMM may be efficient to model a human action.The improved HMM is named Begin-Middle-End Semi-Connected HMM(BME-SCHMM).The main contributions are twofold:(1) each kind of human action is modeled as a non-return three-state HMM;(2) output probabilities in each state are weighted in terms of the key postures position in an action period.Experiments on Weizmann and KTH human action datasets show that compared with the FCHMM based HAR methods,our approach increases the recognition rate,and reduces the computational cost.Therefore,the approach provides a suggestive plausibility proof for the close relationship between human action property and the design of HMM for HAR task.