融合姿态信息和注意力机制的行人重识别研究

Research on person re-identification by fusing posture information and attention mechanisms

针对行人重识别(person re-identification,Re-ID)任务中行人遮挡以及背景信息杂乱不便于提取具有辨识度特征的问题,引入人体关键点模型定位出行人的关键点坐标以便于消除背景信息,根据关键点坐标将图片分割成具有语义信息的区域块。对于骨干网络,为使其提取的特征更加鲁棒,设计一个强化注意力模块(enhanced attention module,EAM),使网络自动分配权重,最终得到更加具有辨识度的特征向量。最后将这些区域块和整体图片送入修改后的注意力机制的神经网络并且联合多个损失一起优化网络。在几个行人重识别数据集试验验证了本研究提出方法优于大多数方法。试验结果还表明该网络针对跨域以及遮挡问题也起到积极作用。

To address the problem of pedestrian occlusion and messy background information in the task of pedestrian re-identification(Re-ID),the human body key point model was adopted to locate the key point information of the pedestrians to eliminate the background information,and the image was segmented into semantic information based on the key point information.In order to make the extracted features of backbone network more robust,enhanced attention module(EAM)was designed,which allows the network to automatically assign weights,and the more recognizable feature vectors were finally obtained.These parts and the overall image were fed into a neural network that incorporates the modified attention mechanism and optimized the network by combining multiple losses.Experiments on several pedestrian re-recognition datasets validate that the proposed method outperforms most state-of-the-art methods.In addition,the experimental results also show that the network has a positive effect on the cross-domain and occlusion problems.