复杂场景下体态深度识别系统架构及核心算法的研究

The posture is the only biological feature that can be non-intrusively captured at a long distance, which makes identification based on the posture draw more and more attention. However, comparing with the exiting methods that are only suitable for simple scenes, the proposed approach will solve the identification by the posture in complex scenarios such cross-view and occlusion. Based on the traditional identification by biological feature, a novel system architecture, combining deep neural network with the tensor model of human posture features, is creatively constructed to deal with identification problem in complicated scenes. Structure of the system is out of the traditional identity recognition technology, fully considering the complex scenario posture characteristics and the characteristics of the actual scene, from strengthening the input information, spatial-temporal information, the characteristics and the space-time status, long-term/short-term time series characteristics, build based on the deep learning of cascade, parallel and extensible depth posture recognition system architecture, to achieve effective extraction combined with various physical characteristics of the degree of differentiation, reduce the complexity of data and so on. And in-depth study on the key point in this architecture, proposed the core algorithm of strengthening judgment network, depth recognition network and deep decision network, effectively raise the universality of the system architecture and the recognition efficiency. Hope that through the objective of the research can promote the depth of the artificial intelligence in the field of public safety applications.

由于人体体态特征具有唯一性和非侵入性，故基于体态特征的身份识别受到越来越多的关注。然而目前的体态识别算法多是针对单一场景，且对真实复杂应用场景中存在的遮挡、跨视角、远距离等问题解决能力有限。本项目拟在当前识别技术的基础上，结合深度学习神经网络技术以及人体特征高阶张量空间建模等方法，创新性地提出复杂场景下基于人体体态特征的身份深度识别的系统架构及其内部核心算法。我们从强化输入信息入手，在充分提取时／空域特征信息的同时，深度挖掘不同特征之间的关系对识别的影响，并建立机制消除数据增量所造成的重复冗余计算，最终形成基于深度学习的级联并行且可扩展的体态深度识别系统架构。并针对该架构中的关键点展开深入研究，拟提出增强判断网络、多维特征高阶张量建模、信息增量计算和深度决策网络等核心算法，切实提高系统架构的普适性及识别效率。希望通过本项目的研究能够促进人工智能在公共安全领域的深度应用。

如何在复杂场景下实现利用体态特征进行身份识别成为深度学习结合体态识别技术的研究热点。本项目在深入研究和分析了深度学习神经网络以及人体特征高阶张量空间建模中的关键技术之后，构建了基于人体体态特征的身份深度识别算法的系统架构模型以及提出一系列核心算法。首先，针对复杂场景下身份深度识别系统架构的设计，我们将传统方法同深度学习神经网络以及张量建模等技术进行有机的融合，采用多级并联的策略，构建了复杂场景下基于体态特征的深度识别系统架构。其次，针对复杂场景下采集图像的真实性和有效性，构建了复杂场景下体态特征增强判断网络模型。针对多维样本特征之间的内部关联性，构建了基于高阶张量人体体态特征的表征模型。最后，针对长时间序列的体态识别，我们还提出了一种基于长时间视频序列特征的深度学习算法。经测试表明，本项目提出的核心算法及系统框架在实际环境中的识别率、错误率等各项指标均较为优异。