高度分层分区的图卷积交警手势识别技术  

Traffic Police Gestures Recognition Based on Graph Convolution with Height Layering Partitioning Strategy

作　　者：张丞 侯义斌 何坚 Zhang Cheng;Hou Yibin;He Jian(Faculty of Information Technology,Beijing University of Technology,Beijing 100124;Beijing Engineering Research Center for IOT Software and Systems,Beijing 100124)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京市物联网软件与系统工程技术研究中心,北京100124

出　　处：《计算机辅助设计与图形学学报》2022年第7期1037-1046,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(52175493);国家重点研发计划(2020YFB2104400).

摘　　要：针对无人驾驶汽车自动识别连续交警手势的需求,提出高度分层分区的图卷积交警手势识别方法.首先,依据人体部件在空间域内的自然、辅助和自身连接关系以及时间域内的关联关系建立交警手势时空图模型,并从图像序列卷积预测模型参数;其次,引入时空图卷积网络,提出以人物自然站立状态下时空图顶点相对高度差为标签的图卷积高度分层分区策略,打破现有分区策略对图结构的限制;最后,设计保留时间维度的空间域平均层网络输出架构,在减少特征数量的同时适配多对多序列预测模式,达到识别连续交警手势的目的.与领域内代表性方法的对比实验表明,该方法的识别准确率显著提高,不同手势之间混淆率仅为0.1%,Jaccard指数超过对比方法.Aiming at the needs for auto-driving vehicles to recognize continuous traffic police gestures,a graph convolutional traffic police gesture recognizer with the height layering partitioning strategy is pro-posed.Firstly,according to the natural,assist,self-connection spatial relationships and temporal associations between human parts,a spatial-temporal traffic gesture model is established.Secondly,the spatial-temporal convolutional network is introduced,and the height layering partitioning strategy is proposed,which uses the relative height differences of parts as labels and breaks the limitations of existing partitioning strategies on graph structure.Finally,the spatial mean layer output structure which retains the length of the temporal dimension is designed to adapt the many-to-many prediction pattern for recognizing continuous traffic police gestures.The experiments show that the proposed method significantly improves the recognizing perform-ance,the confusion rate between gestures is 0.1%and the Jaccard index surpasses comparison methods.

关 键 词：交警手势 连续手势 手势识别 时空图 图卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]