【摘要】 在行人再识别过程中,由于行人姿态变化会导致图像之间对应位置存在身体区域不对齐的问题,从而降低识别准确率。为此,设计一种新的行人再识别方法。利用卷积神经结构计算行人图像的响应图,根据响应图中的极值点定位行人身体节点,并以此划分特征区域,将提取的各个区域的特征进行融合得到特征表示。在比对距离度量上通过引入k倒排近邻使更多的正样本包含在近邻中,在杰卡德距离中将k倒排近邻集编码成向量以减少计算量,使得越近的邻域获得越大的权重。实验结果表明,相比于对整幅行人图像提取特征方法与单独使用马氏距离的方法,该方法能有效提高行人再识别的准确率。更多还原

【Abstract】 In the process of person re-identification,pose variations lead to a problem of misalignment between spatial areas of images,which results in low recognition rate.So this paper proposes a new person re-identification method.Firstly,a convolution neural structure is utilized to calculate responding maps of person images,and the body joints are located according to the extreme points in the responding maps.Secondly,body sub-regions are divided according to the positions of body joints,and they are aligned before further feature extraction.Finally,the features of body sub-regions are fused for identification.By introducing the k-reciprocal nearest neighbor method,more positive samples can be included in the nearest neighbors.With the Jaccard distance,the computation cost is reduced by encoding the k-reciprocal nearest neighbor sets into a vector,which assigns larger weights to closer neighbors.Experimental results show that compared with the feature extraction from the whole person image and using Mahalanobis distance alone,the proposed method can improve the accuracy of person re-identification significantly.更多还原