【作者】 张敏；

【导师】 洪西进（Hong Shi-Jinn）； 李天瑞；

【作者基本信息】 西南交通大学 ， 软件工程， 2022， 硕士

【摘要】 行人检测与重识别工作是计算机视觉领域的研究热点,旨在研究不重叠的多个摄像区域间对于行人的检测以及对特定行人的匹配,多应用于智能城市建设以保障出行安全。但复杂环境下获取的数据往往存在姿势问题、遮挡问题、照明问题、视角问题、背景问题、分辨率问题以及其他开放性问题,本文将针对以上挑战和难点开展研究工作,以促进行人检测与重识别技术在智能城市中的广泛应用,为“平安城市”的建设工作助力。本文基于深度学习技术,以行人检测与重识别研究为主线,主要工作包括以下三个方面:(1)提出一种基于双向特征融合网络的行人检测模型Bi-YOLOv5。首先使用融合CBAM模块的双向特征融合网络增强对行人有效特征的提取能力,然后使用EIo U Loss检测框回归损失函数加快模型收敛,最后将DIoU-NMS作为检测后处理方法进行冗余锚框去除。通过在两个数据集上进行对比实验,证明了所提出的Bi-YOLOv5模型对复杂环境具有更好的兼容性。(2)提出一种基于高阶拓扑关键点特征融合的行人重识别模型HOFRe-ID。首先采用GCN方法对关键点特征建模以获取高阶关系信息,然后在图匹配阶段融合全局特征获取具有更好描述因子的特征,最后使用Batch Hard Triplet Loss指导网络训练以提高模型挖掘困难样本的能力。实验结果表明所提出的HOFRe-ID模型在解决行人遮挡问题具有更好的鲁棒性和有效性。(3)构建行人检测与重识别系统。该系统基于所提出的Bi-YOLOv5和HOFRe-ID模型并使用Qt框架搭建,不受数据获取方式的限制,实现多途径行人数据的检测与重识别,同时使用深度学习方法降低了行人检测与重识别工作的成本,并进一步验证了所提出算法的有效性和实用性。更多还原

【Abstract】 Pedestrian detection and re-identification work is a research hotspot in the field of computer vision.It aims at studying the detection of pedestrians and the matching of specific pedestrians within multiple non-overlapping camera areas,and is widely used for intelligent city’s urban construction to ensure travel safety.However,there exists problems of pose,occlusion,lighting,perspective,background,resolution and other open ones for the data obtained in complex environments.This thesis will focus on the above-mentioned challenges and difficulties to carry out the research work,so as to promote the wide applications of pedestrian detection and re-identification technology in intelligent cities,and support for the construction of “safe city”.Based on the deep learning technology,this thesis focuses on the research of pedestrian detection and re-identification.The main work includes the following three aspects:(1)A pedestrian detection model,named Bi-YOLOv5 with bidirectional-feature fusion network is proposed.Firstly,a bidirectional-feature fusion network with CBAM module is used to enhance the ability to extract effective pedestrian features.Secondly,the EIo U Loss is applied to accelerate the speed of the model convergence.Finally,a detection post-processing method,DIo U-NMS,is employed to remove redundancy of anchor box.Through comparative experiments on two datasets,it proves that the proposed Bi-YOLOv5 model has better compatibility in complex environments.(2)A pedestrian re-identification model,named HOFRe-ID,with feature fusion of highorder topological key points is developed.Firstly,the methods of GCN is used to model key point features to acquire high-order relational information.Secondly,global features are infused in the graph matching phase to acquire better descriptive factors features.Finally,the Batch Hard Triplet Loss is applied to guide network training to improve the model’s ability to mine difficult samples.It shows from the experimental results that the proposed HOFRe-ID model has better robustness and effectiveness in solving pedestrian occlusion.(3)A pedestrian detection and re-identification system is established.The system is based on the proposed Bi-YOLOv5 and HOFRe-ID models and is built by Qt framework.It does not restrict the limitation of data acquisition way,and could realize the detection and reidentification of pedestrian data in multiple ways.In the meantime,the deep learning method is used to reduce the cost of pedestrian detection and re-identification work,and further verify the effectiveness and practicability of the proposed algorithms.更多还原