基于改进CDN的便利店收银台人-物交互检测

Human-Object Interaction Detection of Convenience Store Checkout Counter on Improved CDN

为了提升便利店在收银过程中的规范性和安全性,收集整理了多个在便利店收银台场景下的数据集,通过加入基于HRNet的人体姿势关键点对人-物交互检测方法CDN进行改进,完成了对不同人、物的检测和分类以及人-物之间的交互行为检测,在原有的检测结果上有较大的提升。通过多组对比实验,可以看出增加Transformer编码器内部层数和ResNet结构的复杂度可以获取到更多的图像特征,加入人体姿势关键点能更准确地判断人-物之间是否存在交互关系。此外,将该任务应用实际中,能为便利店的管理提供更加智能化的方法。

In order to improve the standardization and safety of the convenience store in the cashier process,the datasets of multiple convenience store checkout counter are collected and sorted.It improved human-object interaction detection method named CDN by adding the key points of human posture,and finished the detection and classification of different human and objects,which is greatly improved the original detection results.Through multiple sets of comparative experiments,it can be seen that increasing the number of internal layers of the Transformer encoder and the structure complexity of ResNet can obtain more image feature,and adding key points of human posture can more accurately determine whether there is an interaction relationship between humans and objects.In addition,applying this task in practice can provide a more intelligent method for the management of convenience stores.