基于Dirichlet多项式混合模型的复杂人体行为识别

在健康智能照顾护理领域,日常行为识别的准确率至关重要,但是由于日常行为本身的动态可变性以及个体之间的差异性的特点,造成基于可穿戴设备的日常行为识别模型的泛化性差、识别率低,无法对复杂日常行为进行识别的问题。提出一种优化的特征提取方法,将手腕动作聚合为若干个高层语义主题,进而将日常行为表征为一个有序的高层语义主题序列,有效地提升分类的效果。实验结果表明,高层主题语义特征能更准确地表征复杂日常行为的特征,提高了行为识别的准确性。

COMPLEX ACTIVITY RECOGNITION BASED ON DIRICHLET MULTINOMIAL MIXED MODEL

In the healthcare and nursing domain,the accuracy of activity recognition is critical.However,the daily activity is dynamic and varied,and the patterns of the activity vary from person to person,which decreases the accuracy and reliability of activity recognition model and cannot be suitable for identifying the complex behavior.In this work,an optimized feature extraction method is proposed.we improved the algorithm to extract high-level semantic topic features from the hand movements,and then the activity could be represented as a serial of high-level semantic features to improve the practicability and usability.The experiment shows that the system is more suitable for complex activity recognition by extracting high-level sematic features,and it improves the accuracy of recognition.