【作者】 严青；

【导师】 邹腊梅；

【作者基本信息】 华中科技大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 随着监控设备制造技术的成熟,大大小小的公共安全摄像头如雨后春笋般冒出。公共事务大到刑侦破案,小到社区管理都要借助监控找到目标行人。因此对目标行人的定位和检索具有较大的现实意义。行人重识别研究如何跨摄像头找到同一行人。但这个课题的前提是人已经被检测出来,即在事先存在的行人检测图上开展的。因此行人重识别研究有一个现实问题:实际成像的全景图片中不只有行人还有大量的背景,即行人位置的不确定。所以对于实际成像来说,联合行人检测和重识别更加贴合实际。而行人搜索则是将行人检测和重识别联合考虑的课题,具体是指从全景图像中,寻找行人所在的位置并与待查询行人进行匹配。因此本文工作主要围绕着行人搜索展开,具体工作如下:首先,因为摄像头拍摄视角不同,远近不同,得到图片中的行人所占像素大小不同,然而骨干网络往往输出最后一层的特征图进行多任务学习,会导致尺度小的行人特征响应小,难以检出。基于以上问题和相关思考,本文提出了融合浅层、中层和深层特征的特征提取骨干网络,以此为基础在浅层特征上设计空间注意力模块,强化位置信息表达;在深层特征上设计通道注意力模块,强化语义信息的输出。实验表明本文设计的骨干网络能一定程度上提高行人搜索的性能,证明了这种组合的有效性。其次,针对网络空间不变性有限问题,本文设计了一种候选区域精修模块,通过学习仿射变换参数提升网络模型对空间变换的泛化性能。由于行人不是标准矩形,矩形区域采样会不可避免地带来背景噪声。针对背景噪声问题,本文采用一种自适应区域采样模块,该模块对感兴趣区域池化的矩形规则采样进行改进,让每一个采样区域内的小块能够小范围移动,让网络显式地学习并聚焦于行人部分,以此来消除背景噪声带来的影响。实验表明该组合方法能够提升行人搜索性能。最后,从构建细粒度特征表达出发,本文在单步端到端网络框架的后段中设计了一种局部特征和全局特征联合的算法。因为行人特征图从上到下的分块对应人体结构从头到脚的局部部位,所以算法采用对行人特征图进行水平均匀分块的方法来提取局部特征。将多个局部特征和全局特征联合起来作为行人的特征描述,通过引入局部特征这种细粒度信息,丰富行人特征表达。本文进一步提出局部特征筛选模块来消除离群的局部特征点对网络训练产生的扰动。实验表明,联合局部和全局特征的方法能够提升行人搜索的性能。更多还原

【Abstract】 With the maturity of surveillance equipment manufacturing technology,large and small public security cameras have sprung up.From public affairs to criminal investigation and solving cases,to community management,surveillance is required to find target pedestrians.Therefore,it has great practical significance for the location and retrieval of target pedestrians.Person Re-ID studies how to find the same pedestrian across cameras.But the premise of this topic is that people have been detected,that is,carried out on the pre-existing pedestrian detection map.Therefore,there is a practical problem in the research of pedestrian re-identification: not only pedestrians but also a large number of backgrounds in the actual imaged panoramic pictures,that is,the uncertainty of the pedestrian’s position.Therefore,for actual imaging,combined pedestrian detection and re-identification is more practical.Person search is a joint consideration of pedestrian detection and re-identification,which specifically refers to finding the location of pedestrians from panoramic images and matching them with the pedestrians to be queried.Therefore,the work of this thesis mainly focuses on person search.The specific work is as follows:First of all,because the camera’s shooting angle is different,the distance is different,and the pixel size of the pedestrian in the obtained picture is different.However,the backbone network often outputs the feature map of the last layer for multi-task learning,which will result in a small response of pedestrian features with small scales,making it difficult to detect.Based on the above problems and related thinking,this thesis proposes a feature extraction backbone network that integrates shallow,middle and deep features.Based on this,a spatial attention module is designed on the shallow features to strengthen the expression of position information;channels are designed on the deep features.Attention module,which strengthens the output of semantic information.Experiments show that the backbone network designed in this thesis can improve the performance of person search to a certain extent,which proves the effectiveness of this combination.Secondly,in view of the limited invariance of network space,this thesis designs a candidate region refinement module to improve the generalization performance of the network model to spatial transformation by learning affine transformation parameters.Since pedestrians are not standard rectangles,the sampling of rectangular areas will inevitably bring background noise.Aiming at the problem of background noise,this thesis designs an adaptive region sampling module,which improves the rectangular regular sampling of pooled regions of interest,so that small blocks in each sampling region can move in a small range,allowing the network to learn and focus explicitly on the pedestrian part,in order to eliminate the influence of background noise.Experiments show that the combined method can improve the person search performance.Finally,starting from the construction of fine-grained feature representation,this thesis designs an algorithm for combining local features and global features in the latter part of the single-step end-to-end network framework.Because the top-to-bottom blocks of the pedestrian feature map correspond to the local parts of the human body from head to toe,the algorithm uses the method of horizontally evenly dividing the pedestrian feature map to extract local features.Combine multiple local features and global features as the feature description of pedestrians,and enrich the representation of pedestrian features by introducing fine-grained information such as local features.This thesis further proposes a local feature screening module to eliminate the disturbance of outlier local feature points to network training.Experiments show that the method of combining local and global features can improve the performance of person search.更多还原