服务机器人智能空间下目的域无监督人体行为分析

The development of service robots can solve lots of problems caused by the aging population. However, the relatively low intelligent level is the bottleneck for the application of the family service robot. The robot's ability to analyze the human activity restricts its intelligent level. This project solves the unsupervised human activity analysis on the target domain: An analysis method for predicting human activity is presented based on shared semantic space by transfer learning under the intelligent space platform. The research will make the algorithms of human activity analysis meet the needs for theory and application of the service robots. Using the heterogeneous and homogeneous data collected from multiple sources, the project mainly researches the following elements: the co-occurrence mining mechanism based on the homogeneous space-time data constraints and the heterogeneous space-time data constraints on the words of activities, the formation mechanism of the shared semantic space and the human activity analysis algorithm fusing the features and tasks based on shared semantic space. The project will present a human activity analysis model with good reusability and portability. The project will build a preliminary experimental system with the following features: It is supported by the service robot intelligent space. It has good adaptability to the external environment and is effective for the problems with non-labeled samples. Based on this system, the feasibility and the effectiveness of the related theories and techniques will be validated. And the related results will provide important theoretical basis and key technical support for the development and application of the service robots.

发展家庭服务机器人可以解决人口老龄化带来的诸多问题。较低的智能水平是限制家庭服务机器人应用的瓶颈，机器人的人体行为分析能力制约着机器人的服务智能水平。本项目利用服务机器人智能空间平台，借助迁移学习理论，提出一种基于共享语义空间的具有可预测性的人体行为分析方法，以解决目的域下的无监督人体行为分析，使之基本满足服务机器人领域对人体行为分析理论和应用的需要。项目以收集的多源异构数据为基础，重点研究以下内容：基于多源同构和异构数据时间空间上下文约束的共享语义空间生成机制；基于共享语义空间的融合特征层及任务层的知识迁移统一框架。实现服务机器人智能空间支持下的具有良好复用性及可移植性的无监督可预测性人体行为分析系统，并通过搭建的服务机器人智能空间实验系统，验证相关理论和技术的可行性与有效性。项目研究将为服务机器人的发展和推广应用提供重要理论基础和关键技术支撑。

随着我国人口老龄化的加剧，社会劳动力紧缺、空巢老人增多等一系列社会问题不断出现，而解决这些问题的有效途径是发展家庭服务机器人，使机器人替代“人类保姆”走入家庭，为人类服务。而较低的人体行为认知能力是制约现阶段机器人服务智能水平的关键。本项目利用服务机器人智能空间平台，利用知识迁移、深度学习等技术，研究目的域无监督条件下的人体行为分析技术，提升机器人人体行为认知能力。通过深入分析人体行为的时空间关联特点，提出人体行为动静态结合及时空耦合的有效表征方法；针对人体行为复杂性特点，提出基于同时发生约束的多源异构数据跨特征层及任务层的多特征融合方法；针对服务场景下的人体行为数据捕获困难，提出基于知识迁移与注意力机制的目的域无监督/小样本条件下的人体行为分析方法。上述研究大大提升了机器人的人体行为分析能力，在多项行为分析相关竞赛中均取得了优异成绩。项目提出的目的域无监督人体行为分析理论与方法已发表SCI论文14篇，EI论文7篇，国家发明专利授权2项，申请4项。这些理论和方法为人体行为分析、多特征融合等研究提供了新的思路，并促进了机器人理解能力和智能水平的提高，可以应用到服务机器人、助老助残机器人以及其他人机交互领域。

内容获取失败，请点击重试