基于激光扫描和深度学习的人体动作精准识别研究  

Research on accurate recognition of human motion based on laser scanning and deep learning

作　　者：叶裴雷[1] 张大斌 YE Peilei;ZHANG Dabin(Faculty of Megadata and Computing,Guangdong Baiyun University,Guangzhou 510450,China;College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642,China)

机构地区：[1]广东白云学院大数据与计算机学院,广州510450 [2]华南农业大学数学与信息学院,广州510642

出　　处：《激光杂志》2023年第10期178-183,共6页Laser Journal

基　　金：国家自然科学基金面上项目(No.71971089);广东省普通高校重点科研平台(No.2022GCZX009)。

摘　　要：基于特征进行人体动作识别时,主要依靠短时间规模的时空特征,导致识别结果top-1值较低。因此,提出结合激光扫描技术和深度学习原理,设计有效的人体动作识别方法。针对激光扫描数据,进行去噪和校正处理。通过背景差分法划分静态背景与感兴趣区域,根据扫描数据点的连续性特点、空间位置关系,检测出待识别的人体目标。建立包含时空注意力机制的深度学习人体动作识别模型,获取扫描数据中包含的长期、复杂时空信息,得到融合特征输入融合域,得到初始分类结果。最后,结合决策融合机制,生成人体动作精准识别结果。实验结果表明:所提方法的top-1值为74.4%,与其他识别方法相比提升了30.7%、28.9%和24.2%。When human action recognition is performed based on features,it mainly relies on spatiotemporal features on a short time scale,resulting in a low top-1 value of recognition results.Therefore,an effective human action recognition method is proposed by combining laser scanning technology and deep learning principles.Denoise and correct the laser scanning data.The static background and the region of interest are divided by the background difference method,and the human target to be recognized is detected according to the continuity characteristics and spatial position relationship of the scanned data points.A deep learning human motion recognition model including spatiotemporal attention mechanism is established,and the long-term and complex spatiotemporal information contained in the scanning data is obtained.The fusion features are input into the fusion domain,and the initial classification results are obtained.Finally,combined with the decision fusion mechanism,accurate recognition results of human actions are generated.The experimental results show that the top-1 value of the proposed method is 74.4%,which is increased by 30.7%,28.9%and 24.2%compared with other recognition methods.

关 键 词：激光扫描 深度学习 人体动作识别 时空金字塔 注意力机制 目标检测 

分 类 号：TN911[电子电信—通信与信息系统]