基于深度卷积递归神经网络的多视角步态识别研究

Gait recognition is a new biometrics technique, and it is the only way to identify a human remotely compared with other biometric techniques. However, pedestrians will be affected by the external environment and its own factors, which lead to gait features show a strong intraclass changes. The change of viewing angles or the directions of walking is one of the most important factors influencing the performance of gait recognition system. Aiming at the difficulties and pain points of gait recognition system in real monitoring environment: the effective representation of the two-dimensional gait sequence signals under the premise of not losing the time domain information; The gait view needs to be estimated in advance and the estimation inaccurate will affect the recognition performance; The contradictions between the cost constraints of gait registration and gait recognition accuracy under the multi-view (including overlooking) situation. This project proposes a multi-view convolution recurrent neural network for effective representation of the two-dimensional gait sequence; this project propose a deep generating model to generate the multi-view (Or multi-state) gait sequences, which improves the performance of multi-view gait recognition while reducing the registration cost of gait and improving the registration convenience; this project propose a multi-view convolutional LSTM model which does not need the viewing angle estimation to solve the problem of cross-view gait similarity measure in the recognition process through a deep network. An end-to-end deep multi-task network is proposed to efficiently perform all the procedures in the traditional gait recognition, such as pedestrian detection, foreground extraction, gait silhouettes normalization, multi-view gait feature extraction, and identification, in one deep neural network. The research results will provide new methods and new ideas for gait recognition technology.

步态识别是一种新兴的生物特征识别技术，与其他生物特征识别技术相比，步态识别是唯一可以远距离识别的方法。针对实际监控环境下步态识别系统存在的难点：不丢失时域信息前提下步态序列的有效表示问题；待识别步态视角需要预先估计，估计不准确会影响识别性能问题；步态注册的成本限制和多视角（包括俯视）步态识别准确率的矛盾。本项目提出多视角卷积递归神经网络对两维步态序列信号进行有效表示；提出生成多视角（或多状态）步态序列的深度生成模型，在降低步态注册成本、提高注册便捷性的同时，增进多视角步态识别的性能；提出基于多视角卷积LSTM的模型，不需要视角预估计，通过一个深度网络解决识别过程中的跨视角步态相似性度量问题；提出一个端到端的深度多任务网络，高效完成传统步态识别的行人检测、前景提取、步态轮廓归一化、多视角步态特征提取、识别等所有步骤。研究成果将为步态识别技术提供新方法、新思路。

项目开展了基于深度神经网络的多视角步态识别研究工作。多视角步态识别在各地机场、海关等出入境口岸的安检系统；各大银行、金库等具有防盗反恐需求的安全认证系统，各大重点实验室和涉密部门的智能监控系统；各大医院的医疗诊断系统；以及全国公安部门的刑事侦查系统存在广泛的应用前景。. 本项目的研究提出了基于共享表示的多视图深度生成网络，从生成模型的角度构建视角转化模型，并将其扩展成动态生成模型以处理多视图时间序列数据，解决了基于步态能量图和步态视频序列的多视角步态识别，异质特征度量和融合，缺失数据补齐和数据增广等相关问题。. 提出了形变生成网络模型，以纯无监督的形式解耦分离表观和几何信息。该模型可以解耦分离数据的表观和几何信息，解耦后的几何信息可以迁移到其他的数据集合，并提高识别的性能。扩展的动态形变生成模型可以用于处理视频序列数据，提取的动态形变场信息编码了目标的动态几何形变特征，独立于目标的表观信息，可以应用于步态识别、动作识别、人脸表情识别等本质上由动态几何信息决定分类结果而又容易受到表观信息影响的应用上。. 提出了基于多尺度特征集融合的端到端孪生网络的多视角步态识别，并初步研究了稀疏化生成神经网络模型诱导的可解释层级复合模型，将两个重要的top-down模型，深度生成模型和稀疏表示模型无缝结合在统一的概率框架下。通过引入“交叉熵三角”目标函数，将最大似然学习和对抗学习及统计物理上的标准能量函数无缝结合在一起。提出的模型可以学习到“场景-目标-部分-子部分-原语”等一系列层级复合关系，学习到的解析语法树具有可解释性。初步研究了具备可解释性的步态识别技术，研究人体的哪些部位的几何姿态信息及动态信息最有利于提升步态识别的准确率，使步态识别具备可解释性。利用现有的提取人体骨骼关键点的深度学习模型，提取关键特征点信息，然后通过优选特征关键点，找出最有利于步态识别的特征点，使得步态识别过程具备可解释性。