基于强化深度特征融合的行人再识别算法研究

Pedestrain re-identification based on reinforced deep feature fusion

行人再识别指的是在非重叠的多摄像头下匹配行人目标,通常由特征表示和度量学习两个部分组成。文中针对特征表示进行研究,提出一种新的行人再识别特征提取算法,在深度学习特征的基础上加入了基于人体结构检测的多分类特征,建立了强化深度特征的多特征融合模型。经过对深度学习特征的强化训练,文中得到一种描述性更强、更有效的融合特征,并且在多个公开数据集上有很高的识别率。

Pedestrain re-identification refers to the matching of pedestrian targets in non-overlapping multi-cameras,which is usually composed of feature representation and metric learning. This paper focuses on the former,and proposes a novel feature extraction algorithm of pedestrain re-identification.By adding multi-classification features based on human body structure to the convolutional neural network（CNN） features,it proposes a reinforced deep feature fusion algorithm. Through the reinforcement of CNN features,it obtains a more descriptive and effective fused feature,and gets high recognition rate on many public data sets.