基于fHMM分类优化的多传感器手语手势识别方法

Optimized Strategy of fHMM for Real-time Multi-sensor Gesture Recognition

目的探索基于角速度、加速度、表面肌电信息融合的算法,使在嵌入式设备上实现实时手势识别成为可能。方法以表面肌电样本熵检测动作起止点,配合多级决策树融合轨迹和肌电信息实现手语手势的识别;采用分帧隐马尔可夫(framing Hidden Markov Model,fHMM)优化策略降低识别延时;在运行速度为300 MHz的嵌入式软件上进行算法测试。结果融合三类信息后,30个中国手语词获得97.5%±1.6%的识别率,角速度的加入使得识别率平均提高4%;同时,使用基于f HMM的分类优化策略将平均识别延时降低至(175±38)ms,减小约670 ms。结论本文为实时手语手势识别设备的研制提供了一种可行的方案。

Objective To explore the feasibility of Chinese Sign Language （CSL） recognition based on acceler- ometer （ACE）, angular velocity （AV） and surface electromyogram （sEMG） and minimize the identification delay on the basis of ensuring the recognition rate to make the algorithm running in real-time system. Methods The sample entropy was proposed to detect sign word segments within a sequence of signals. And then a hier- archical decision tree was constructed for the information fusion of ACC, GYR and sEMG signals to realize rec- ognition of CSL. An optimized strategy of framing Hidden Markov Model （fHMM） was proposed to reduce i- dentification delay. The algorithm was performed on a dedicated embedded software with 300 MHz. Results The average recognition accuracies of 30 CSL was 97.5% , improved by 4% , after using AV. The average i- dentification delay was 175 ms, reduced by 670 ms, after using the optimized fHMM. Conclusion This meth- od provides a feasible way for realizing real-time gesture recognition system.