基于自适应关键点融合的人员穿戴属性度量分析方法研究

Research on Metric Analysis Method of Personnel Wear Attribute Based on Adaptive Key Point Fusion

大多数人员穿戴属性研究使用图像整体特征构建多分类问题,存在空间信息耦合和对穿戴款式扩展支持不好的问题。提出了一种基于自适应关键点融合的人员穿戴属性度量分析方法。该方法由特征提取网络和特征向量度量学习结构两部分组成,特征提取网络包含3部分模块:基于HRNet网络的行人关键点热力特征图模块、自适应关键点与图像特征融合模块、带注意力机制的残差网络特征映射模块。基于该网络的自适应机制实现头部、上衣、下装3个属性识别分支的特征向量的映射。针对每个属性识别分支穿戴的不同状态或款型,利用基于arcface的度量学习方法,实现穿戴属性识别。结果表明,提出的方法在精细标注的穿戴款式数据集上识别精度为94.2%。在某实际场景工装数据集上,仅通过训练集入库,工装穿戴识别精度达到86.6%。研究可为人员穿戴状态和款型分析提供参考。

Most studies regard the wear status of personnel as a classification problem,and the model is poor at recognizing emerging wearables.In this paper,the safety protection wear status is taken as a person attribute,and then a human wearable recognition algorithm based on adaptive component segmentation is proposed.The method consists of two parts,feature extraction network and metric learning.The feature extraction network consists of three modules,namely,pedestrian key point thermal feature map module,adaptive key point and image feature fusion module based on HRNet network,residual network feature mapping module with attention mechanism.Based on the adaptive mechanism of the network,the feature vector mapping of the three attribute recognition branches including head,top and bottom is realized.Using the measurement learning method based on arcface,the different states or models of the branch wear are recognized for each attribute.The results show that the method proposed in this paper has a recognition accuracy of 94.2%on the data set with finely marked wear styles.On a practical scene tooling style dataset,we only put the training set into the database,so the accuracy of tooling wearing recognition in the test set reached 86.6%.Our method can provide a reference for the analysis of personnel wear status and model.