面向可穿戴用户行为识别的增量学习方法研究

Wearable activity recognition perceives users’ activities by wearable devices, which is widely adopted in different fields, such as health care, sports health, industrial production, entertainment, etc. However, affected by the elements such as the dynamic of data, diversity of activity category, personalized user activity and difference of application requirements, the commonly adopted fixed activity recognition model cannot satisfy the demand of accuracy and practicability. Therefore, this project studies the incremental learning method for wearable activity recognition based on large-scale data, to construct a unified data-feature-classification incremental learning model. This project includes: 1) Fusion technology for the large-scale heterogeneous collaborative data. Construct a large-scale open data set by the fusion of heterogeneous data to extract the optimal features; 2) Incremental learning model for wearable activity recognition. Study the unified incremental learning model to implement the update at data level, feature level and classification level; 3) A rapid assessment method. Build a test platform to rapidly assess the recognition performance, resource consumption, user experience of the learning model, and verify the research results in all aspects.

可穿戴行为识别通过用户穿戴传感设备来感知其行为，目前广泛的应用于健康监护、运动保健、工业生产及娱乐等领域。然而，受数据动态性、行为类别多样性、用户行为个性化和应用需求差异性等因素的综合影响，目前常采用的固定行为识别模型往往在精度和实用性等方面都难以满足需求。本项目研究可穿戴用户行为识别的增量学习方法，在大规模行为感知数据的基础上，构建融合数据、特征和类别的统一增量学习模型。研究内容包括：1）大规模异构协同数据的感知融合技术。构建大规模开放数据集，通过异构数据的感知融合提取最优特征集，对用户行为进行表征；2）面向可穿戴用户行为识别的增量学习模型。研究统一的可穿戴行为识别增量学习模型，同时实现数据层、特征层和类别层的增量更新；3）可穿戴用户行为识别模型的快速评估方法。搭建实验测试平台，对模型的性能、资源消耗以及用户体验等进行快速评估，对研究成果进行全方位验证。

受数据动态性、行为类别多样性、用户行为个性化和应用需求差异性等因素的综合影响，目前常用的固定可穿戴行为识别模型往往在精度和实用性等方面都难以满足需求。针对这个问题，本项目对面向可穿戴用户行为识别的增量学习方法展开研究。研究实现了大规模异构协同数据的感知融合方法、面向可穿戴行为识别的增量学习模型以及可穿戴用户行为识别模型的快速评估方法。相关成果在IEEE Transactions on Knowledge and Data Engineering (IEEE TKDE)、ACM International Conference on Multimedia (ACM MM)、ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp)等国内外知名期刊会议上发表和接收论文34篇，申请国家发明专利及实用新型专利20项、软件著作权2项。共参加国际学术会议7人次，邀请国内外知名学者交流访问9人次。培养8名博士和8名硕士。获得2017年CCF科学技术奖技术发明一等奖等奖励。