基于人体多部位传感器和穿戴式视频的动作识别研究

Human activity recognition is a key technology of human-computer interaction and healthcare. Current activity recognition methods based on wearable device raise the low recognition accuracy and time-consuming problems In order to solve these problems, we investigate combining motion sensor data, Global Positioning System (GPS) data and video data from multi-position wearable devices for activity recognition. First, due to the low recognition accuracy of current activity recognition methods based on data of sensors, multi-position sensors are utilized for data capture and a hierarchical neural network method for activity recognition based on sensor data is proposed to improve the accuracy of classification. Secondly, we investigate the feature extraction method based on wearable video. Since current image feature extraction methods for activity recognition didn’t use static scene information, a new feature which combines motion feature and static feature is proposed to improve the accuracy. Finally, for the time-consuming problem of multi-source data, an adaptive Hidden Markov Model is proposed. High computing speed nodes are imported to the model firstly, which could avoid the problem of redundancy. With the help of this method, the speed of the classification is increased. A human activity recognition prototype system under the real envrionment will be constructed, which provides technical support for human-computer interaction and healthcare.

人体动作识别是人机交互、健康监护等多种应用的关键技术。本课题针对现有基于穿戴式设备的动作识别方法准确率低、时间复杂度高的问题，拟通过融合多部位穿戴式设备所获得的运动传感器数据、全球定位系统数据及视频数据等多源数据开展深入研究：首先，针对现有基于传感器数据的动作识别方法分类精度较低的问题，拟采用人体多部位传感器进行数据采集，并利用分层的神经网络方法进行动作分类，以提高识别精度；其次，拟对基于穿戴式视频数据的特征提取方法进行研究，由于现有的面向动作识别的视频特征缺少对静态场景的描述，因此本课题拟融合动态特征与静态特征进行动作识别，提高识别准确率；最后，针对多源数据导致的运算量过大的问题，拟研究自适应隐马尔可夫模型，优先选择特征提取速度快的传感器进行识别，避免冗余信息对于算法运行时间的影响，提高识别速度。最终搭建真实环境下的人体动作识别原型系统，为人机交互及人体健康监护等领域提供技术支持。

人体动作识别是人机交互、健康监护等多种应用的关键技术。本项目研究了基于可穿戴设备运动传感器数据、视频图像数据以及上下文数据的人体动作识别方法。首先，研究了基于运动传感器的动作识别方法，提出了基于穿戴设备的上下文环境融合分层动作识别方法，提出了基于智能可穿戴手表的实时动作识别和运动计数方法来评估运动量，而无需针对各项动作单独设计评估算法。其次，研究了基于视频图像数据的动作识别方法，总结了基于视觉的人体动作识的最新研究进展，提出了基于彩色和深度图像传感器的人体摔倒自动检测方法。最后，研究了基于数据融合的动作识别方法，构建了融合智能手表和彩色深度摄像机的运动捕获系统，提出了多源数据映射方法，为了有效地利用所有数据源，提出了基于多源数据的分组选择方法，自动寻找最佳特征组合并建立分层识别结构，有效的提升了人体动作识别的准确性。.本项目已发表学术论文10篇，其中SCI国际期刊论文5篇。本项目研究成果在公共安全，智能监控等、运动监控领域具有广泛的应用场景。