多视图深度学习的RGBD人体行为识别与理解

The project aims at studying the relevant theories and techniques about multi-view deep learning based RGBD human action recognition. Specifically, multiple stacked Auto-Encoders are firstly used to learn the compact yet low-dimensional feature representations from each modality. Then, a multi-view feature fusion technique is adopted to optimally fuse these multi-modal features together. Finally, with the aid of appropriate deep learning and classification/recognition algorithms, the goal of fast RGBD human action recognition and understanding can be achieved. The research contents include several key theories and techniques such as high-definition depth image data generation, high-quality 3D human motion data reconstruction, multi-view deep learning based discriminative visual feature generation and deep learning framework based RGBD human action recognition. The research outcomes will be the design and implementation of high effective and robust algorithms, as well as forming a multi-view deep learning based RGBD human action recognition prototype system. Besides, this project will publish about 6 to 8 high-level academic papers in the domestic and foreign important journals and conferences, and apply for 3 to 5 national patents and software copyrights, and train 6 to 10 PhD and Master students.

本课题以基于多视图深度学习的RGBD人体行为识别相关理论与技术为研究目标，采用多个叠加的自动编码器模型分别从不同模态数据中学习得到低维紧凑的特征表达，利用多视图特征融合技术对多模态特征进行最优化融合，并结合使用相应的深度学习与分类识别算法，最终实现快速RGBD人体行为识别与理解。具体研究内容包括高质量深度图像数据生成、高质量三维人体运动数据重建、基于多视图深度学习的鉴别性视觉特征生成以及基于深度学习框架的RGBD人体行为识别等关键理论方法与技术。研究成果将体现为高效、鲁棒算法的构造和实现，并形成一个基于多视图深度学习的RGBD人体行为识别原型系统，在国内外重要学术期刊和会议上发表高水平学术论文6-8篇，申请国家发明专利与软件著作权3-5项，培养博士、硕士研究生6-10人。

本课题以基于多视图深度学习的RGBD人体行为识别相关理论与技术为研究目标，采用多个叠加的自动编码器模型分别从不同模态数据中学习得到低维紧凑的特征表达，利用多视图特征融合技术对多模态特征进行最优化融合，并结合使用相应的深度学习与分类识别算法，最终实现快速RGBD人体行为识别与理解。具体研究内容包括高质量深度图像数据生成、高质量三维人体运动数据重建、基于多视图深度学习的鉴别性视觉特征生成以及基于深度学习框架的RGBD人体行为识别等关键理论方法与技术。研究成果为高效、鲁棒算法的构造和实现，在国内外重要学术期刊和会议上发表高水平学术论文13篇，申请国家发明专利3项，软件著作权2项，培养博士、硕士研究生6人。