基于深度学习的多特征融合人体动作识别算法

徐海宁1; 王彦坤2; 3; 樊勇4; 文奴2

基于深度学习的多特征融合人体动作识别算法

引用本文：徐海宁1，王彦坤2,3，樊勇4，文奴2.基于深度学习的多特征融合人体动作识别算法[J].计算技术与自动化,2025,(4):31-37

摘要点击次数: 8

全文下载次数: 2

作者	单位
徐海宁1，王彦坤2,3，樊勇4，文奴2	（1.中国通信建设集团设计院有限公司数字化咨询中心，河南郑州 450052；2.深圳职业技术大学物联网研究院，广东深圳 518055 3. 深圳大学建筑与城市规划学院智慧城市研究院，广东深圳 518061；4.深圳职业技术大学人工智能学院，广东深圳 518055）

中文摘要:针对传统特征提取方式计算量大，识别率低等问题，提出了一种基于深度学习的多特征融合人体动作识别算法。方法首先提取原始深度视频序列，接着对每帧图像旋转特定角度补充更多视角下的动作信息，提出自适应时间模型将深度视频序列划分为若干子动作。通过累积每帧图像之间能量变化较大的部分形成动态能量图，累积能量变化较小的部分形成静态能量图，统一称之为运动能量图。引入多通道卷积神经网络提取运动能量图中的动态特征和静态特征；引入自适应矩估计调整神经网络训练中每个参数的学习率，提高模型训练的效率和稳定性；引入L2范数正则化，减小模型复杂度并防止过拟合。在公开的数据集UTD-MHAD、MSR Action3D和MSR DailyActivity3D数据库上测试，算法能够有效地提取动作的运动信息并取得较好的识别结果，与主流方法比有一定的提升。

中文关键词:自适应时间模型运动能量图多通道卷积神经网络自适应矩估计 L2正则化

Multi-feature Fusion Algorithm for Human Action Recognition Based on Deep Learning

Abstract:To address the issues of high computational complexity and low recognition rates associated with traditional feature extraction methods, a deep learning-based multi-feature fusion algorithm for human action recognition is proposed. The method begins by extracting the original depth video sequence, followed by rotating each frame image by a specific angle to supplement action information from multiple viewpoints. An adaptive temporal model is proposed to divide the depth video sequence into several sub-actions. Dynamic energy maps are formed by accumulating portions with relatively larger energy changes between each frame of image, while static energy maps are formed by accumulating portions with relatively little energy changes. Both are collectively referred to as motion energy maps. A multi-channel convolutional neural network is introduced to extract dynamic and static features from the motion energy maps. Adaptive Moment Estimation is used to adjust the learning rate of each parameter during neural network training, improving the efficiency and stability of model training. L2 norm regularization is introduced to reduce model complexity and prevent overfitting. Tested on the public datasets of UTD-MHAD, MSR Action3D and MSR DailyActivity3D, the algorithm can effectively extract the motion information of actions and achieve good recognition results, which is a certain improvement compared with the mainstream methods.

keywords:adaptive temporal model motion energy map multi-channel convolutional neural network adaptive moment estimation L2 regularization

查看全文 查看/发表评论 下载pdf阅读器