结合全局上下文信息的交警手势识别方法  

Traffic police gesture recognition method combined with global context information

作　　者：程贝芝 伍鹏[1] 寇静雯 何一鸣 谢凯[1,3] 盛冠群 CHENG Beizhi;WU Peng;KOU Jingwen;HE Yiming;XIE Kai;SHENG Guanqun(School of Electronic Information,Yangtze University,Jingzhou 434023,China;School of Computer and Electronic Engineering,Guangxi University,Nanning 530000,China;Western Research Institute of Yangtze University,Karamay 834000,China;School of Computer and Information,Three Gorges University,Yichang 443002,China)

机构地区：[1]长江大学电子信息学院,荆州434023 [2]广西大学计算机与电子信息学院,南宁530000 [3]长江大学西部研究院,克拉玛依834000 [4]三峡大学计算机与信息学院,宜昌443002

出　　处：《中南民族大学学报（自然科学版）》2023年第3期349-356,共8页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家自然科学基金项目(42204111);新疆维吾尔自治区自然科学基金项目(2020D01A131);长江大学大学生创新创业训练计划项目(Yz2021040)。

摘　　要：为了解决现有交警手势识别模型对全局信息提取得不够完善而导致识别准确率不高的问题,提出了一种结合全局上下文信息的交警手势识别方法(Spatial Graph Convolutional and Temporal Transformer network,SGCTT).该算法以STGCN网络模型为基础,使用图卷积进行非线性映射的方式来实现空间编码,并将空间编码得到的高维特征合并上一个用于分类的维度相同的可训练向量,然后再加入时间位置信息进行时间编码.最后,将时间编码器输出的训练完成的向量通过一个多层感知机进行分类输出.实验结果表明:在光照状态良好和夜间灯光下,SGCTT算法满足实时性的要求,对交警手势检测的平均精度达到了97.83%,相比于现有的交警手势分类模型有所提高,且具有更强的可移植性.In order to solve the problem that the current traffic police gesture recognition model does not extract the global information perfectly,which leads to the low recognition accuracy,a traffic police gesture recognition method combining the global context(SGCTT)is proposed.The algorithm is based on STGCN network model,and uses graph convolution to realize spatial coding by nonlinear mapping.The high-dimensional features obtained from spatial coding are merged into a training vector with the same dimension for classification,and then the time position information is added for temporal coding.Finally,the training vectors output by the time encoder are classified and output by a multi-layer perceptron.The experimental results show that the SGCTT algorithm meets the real-time requirements under good lighting conditions and night lights,and the average accuracy of the traffic police gesture detection reaches 97.83%,which is improved compared with the existing traffic police gesture classification models,and has stronger portability.

关 键 词：交警手势 全局信息 连续手势 行为识别 姿态估计 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]