基于HMM的动作识别结果可信度计算方法

Identifying the confidence level of activity recognition via HMM

针对当前动作识别可信度计算方法中混淆率高、不适用于迁移学习等问题,提出一种基于样本上下文信息的可信度计算方法(S-HMM,sliding windows hidden Markov model)。该方法使用隐马尔可夫模型(HMM,hidden Markov model)理论对识别结果序列建模,将样本所在序列识别正确的概率作为识别结果的可信度,避免了当前可信度计算方法依赖于样本在特征空间中分布的问题。实验使用真实场景中的数据进行仿真,结果表明,与现有方法相比,该方法可将可信度混淆率降低37%左右。

A context-based method to identify the confidence level of activity recognition was proposed, referred to as S-HMM（sliding window hidden Markov model）, which reduced the confusion rate and facilitated the transfer learning. With S-HMM, the activity recognition sequence was modeled as HMM（hidden Markov model）and the corresponding probability was adopted as the confidence level. This way, S-HMM removed the dependency of the confidence level on the sample distribution in the feature space. S-HMM is extensively evaluated based on real-life activity data, demonstrating a reduced confusion rate of 37% when compared to the state-of-the-art methods.