基于深度学习的复杂场景行人再识别若干关键问题研究

The variations including color, illumination, posture, imaging quality as well as low-resolution of the captured frames cause large appearance variance across multiple cameras, which make person re-identification still a challenging problem. Most of the existing researches extract handcrafted features or deep features from intact cropped still pedestrian images to perform re-identification, which are difficult to adapt the severe occlusion and deformation of pedestrian images in complex scenes. This project stimulates the visual process of human and adopts deep learning framework to solve the person re-identification problem in complex scenes. This project mainly focus on: 1) the still image-based person re-identification utilizing the local region feature of pedestrian image to improve the matching rate; 2) the video-based person re-identification exploiting the appearance information and motion information of video frames to solve the problem of low matching rate in single frame image-based method; 3) person search in complex scenes, integrating the pedestrian detection and re-identification to improve the pedestrian matching rate. This research plays an important role in the development of artificial intelligence and the intelligent video surveillance system.

由于图像分辨率差异、光照差异、行人姿态差异以及摄像机视角和成像质量差异等原因, 导致同一行人在不同监控视频中的外观区别很大，给行人再识别带来了巨大挑战。现有研究大多数是利用手工设计特征或者深度特征对分割完整的静态行人图象进行行人再识别，不适合行人严重遮挡、变形等复杂场景。本项目模拟人的视觉辨识过程，采用深度学习框架实现复杂场景行人再识别。项目主要内容为：（1）研究复杂场景下静态图片行人再识别问题，利用行人图片的局部特征提高识别率；（2）研究视频行人再识别问题，综合利用视频图像提供的外观特征和运动特征，解决单帧图片识别率低的问题；（3）研究复杂场景行人搜索问题，将行人检测和识别统一起来，提高复杂场景下行人识别的准确率。项目研究成果对促进人工智能学科发展、提高视频监控系统的智能化程度具有重要作用。

由于图像分辨率差异、光照差异、行人姿态差异以及摄像机视角和成像质量差异等原因，导致同一行人在不同监控视频中的外观区别很大，给行人再识别带来了巨大挑战。现有研究大多数是利用手工设计特征或者深度特征对分割完整的静态行人图象进行行人再识别，不适合行人严重遮挡、变形等复杂场景。本项目模拟人的视觉辨识过程，采用深度学习框架实现复杂场景行人再识别。项目主要内容为：（1）研究了复杂场景下静态图片行人再识别问题，提出了多种行人静态图片的局部自动划分算法，有效提升了存在背景干扰、遮挡场景下的行人特征判别力。（2）研究了视频行人再识别问题，提出自适应学习视频中行人时间-空间特征的端到端深度模型，通过结合多帧图片的外观特征和运动特征，解决了单帧图片识别率低的问题。（3）研究了复杂场景行人搜索问题，提出了一种边查找边匹配的行人搜索框架，将提高复杂场景下行人搜索的准确率。项目研究成果将用于开发行人再识别应用演示系统，提高了视频监控系统的智能性。