基于深度学习的步态识别研究

Gait recognition aims to recognize a person’s identity by studying human walking features. Compared with fingerprint recognition and face recognition, gait recognition owns advantages of long-distance, non-contact and hard-camouflage, and is attracting more and more attention. In recent years, deep learning methods have presented great potential for improving gait recognition accuracy, but lack of public training samples is a major obstacle. Since gait is easily influenced by several interference factors, such as views, clothes, carryings and so on, it is very difficult to collect gait samples containing all interference factors. This project proposes a gait video synthesis method based on deep networks, which can automatically generate gait videos under different views and with different clothes. This project generates samples under different views by performing view projection on 3D poses of target samples, projects real and synthetic samples to common spaces to reduce the influence of distribution differences, and translates clothing region masks to reduce clothing disturbances. This project realizes gait dataset expansion saving a lot of labor cost, which can provide adequate training samples for gait recognition methods and improve gait recognition performances to new levels.

步态识别是指通过研究人行走的步态特征对其身份进行识别。识别样本可以通过远程拍摄的方式较为容易地获得，相比传统的指纹、人脸识别，步态识别具有远距离性、非接触性、难以伪装性等优点，正在受到越来越广泛的关注。近年来深度学习的迅速发展对步态识别的精度有较大的提升，但公开的训练样本不足仍是较大的制约。由于步态极易受到视角、服饰等影响因素的干扰，采集到所有干扰类型下的大量样本非常困难。本项目提出基于深度学习的步态视频合成方法，可自动合成在不同视角、服饰下的步态视频，以增加步态数据库的样本数量，提高识别精度。本项目通过对目标样本进行3D姿态投影降维，合成不同视角下的样本，将真实样本与合成样本映射到共同空间中以减少分布差异的影响，并通过服饰区域掩膜转换降低服饰干扰。本项目在节省大量人力成本的情况下完成步态数据库的扩展，可使步态识别精度上升到新的台阶。

步态识别是指通过研究人行走的步态特征对其身份进行识别。识别样本可以通过远程拍摄的方式较为容易地获得，相比传统的指纹、人脸识别，步态识别具有远距离性、非接触性、难以伪装性等优点，正在受到越来越广泛的关注。近年来深度学习的迅速发展对步态识别的精度有较大的提升，但公开的训练样本不足仍是较大的制约。由于步态极易受到视角、服饰等影响因素的干扰，采集到所有干扰类型下的大量样本非常困难。本项目提出基于星形生成对抗网络的步态样本合成方法，可自动合成在不同视角、服饰下的步态样本，且提出将真实样本与合成样本映射到共同空间中以减少分布差异的影响，并通过标签编码转换提高不同服饰条件下的步态样本合成质量。本项目的研究成果表明，通过设计生成对抗网络模型合成步态假样本，可以在节省大量人力成本的情况下完成步态数据库的扩展，并可使步态识别精度显著提升。