基于回归的视角转换框架下的多视角行人步态识别

As a biometric technique, human gait have been used to identify people, particularly, at a long distance when other techniques (e.g., face recognition and fingerprint recognition) fail. Yet, to date, the challenge of how a computer can uniquely recognize human gaits when they are recorded by cameras under different views remains. Specifically, the main objectives of this project can be divided into following parts: develop robust cross-view gait recognition under large view change; recognize gaits when the viewing angle continuously changes and recognize gait under the view angle on which there is no prior knowledge. This research is based on our preliminary work on regression framework which has achieved encouraging progress. To address these challenges, this research will focuses on: 1) develop a new framework for building a new view transformation model for accurate and robust gait recognition under large cross views; 2) conduct gait recognition when viewing angle continuously changes by taking multiple views as a whole; 3) establish many-to-one mapping strategy for predicting gaits at a target view. ?This project will significantly boost the recognition accuracy of current gait recognition systems.

作为一种特有的生物识别技术，行人步态可以用来识别远距离、影像质量不高或图像模糊的人体目标,可有效补偿人脸识别、指纹识别等技术在实施时的局限，对保障和维护社会公共安全有重要意义。然而目前步态识别技术在大视角变化、视角连续变化等多视角情况下的效果仍亟待提高。基于我们前期在步态识别取得的成果，本课题拟利用回归框架来处理上述问题，主要技术措施为：1）利用回归框架来建立视角转换模型，使之能够准确且鲁棒的处理大视角变化情况下的步态；2）将多个视角的步态特征作为整体进行分析来解决视角连续变化的情况；3） 针对无视角先验知识的情况，建立一个多对一的映射策略来预测单目标视角下的步态。预期本课题将有效的提高现有多视角行人步态识别系统的识别率。

鉴于行人步态可以在远距离目标识别中发挥重要作用，步态识别技术正受到越来越多的研究人员的关注。然而步态识别技术的准确性和鲁棒性受到诸多因素的影响，尤其是在视角变化和衣着变化等情况下的效果亟待改善。本课题采取了多种手段来提高步态识别的鲁棒性和准确性，主要技术措施为：1）通过多场景下的动态特征提取技术，准确且鲁棒的实现了视角变化情况下的步态识别；2）通过构建基于人体模型的步态特征，准确且鲁棒的实现了视角变化和衣着变化情况下的步态识别；3）借助于深度学习框架，将多个视角的步态特征作为整体进行分析，解决了大视角变化和视角连续变化的情况；4） 针对无视角先验知识的情况，实现了基于步态特征的行人再识别。