【作者】 李慧；

【导师】 裴小兵；

【作者基本信息】 华中科技大学 ， 软件工程， 2021， 硕士

【摘要】 人体姿态估计是计算机视觉领域中极具挑战性的问题之一,目标是检测出图像数据中人体的骨骼关节点。作为辅助计算机分析目标活动和行为的关键技术,不仅决定着人体动作识别的效果,还在促进智能机器人、VR以及人机交互等领域的发展担任着重要的角色。同时,在视频监控和健身指导中,人体姿态估计也有着普遍的应用。为人类生活提供了极大的便利。近年来,人体姿态估计的算法性能得到飞跃式的提升,为家居环境下的摔倒检测提供了有力的技术支持。意外摔倒威胁着人们尤其是独居老人的健康生活。为了解决这个问题,提出面向家居场景的人体姿态估计系统。主要实现了姿态估计、摔倒检测、监控预警和信息管理等功能。选取Center Net算法实现人体姿态估计,针对Center Net在家居场景下出现关节点漏检和误检的情况,改进模型的卷积算子,并收集家居场景下的图片,通过对数据标注、标签转换以及与COCO数据集的合并,构建了一个数据量充足、场景丰富的训练集。改进后的算法能有效提升原Center Net模型的精度。提高在家居场景中的应用效果。用于摔倒检测具有较好的鲁棒性。运用姿态估计算法获取人体骨架信息,通过计算人体宽高比、高度变化率、躯干倾斜角三个姿态特征对目标进行摔倒判别。检测到摔倒行为时发送预警信息,以提供目标对象及时的帮助和救治。完成系统的需求分析后,构建整体架构,对系统的各个功能模块进行划分和设计,并使用Py Qt5框架实现了面向家居场景的人体姿态估计系统。测试系统的功能和性能,具有良好的稳定性和扩展性。面向家居场景的姿态估计系统在真实家居场景下测试结果具有高准确性和实时性。具有一定的实用价值。更多还原

【Abstract】 Human pose estimation is one of the challenging problems in the field of computer vision,where the goal is to detect skeletal joint points of the human body in image data.As a key technology to assist computers in analyzing target activities and behaviors,it not only determines the effectiveness of human motion recognition,but also plays an important role in facilitating the development of intelligent robots,VR,and human-computer interaction.Meanwhile,human pose estimation also has common applications in video surveillance and fitness instruction.It provides great convenience for human life.In recent years,the algorithm performance of human posture estimation has been improved by leaps and bounds,providing powerful technical support for fall detection in the home environment.Accidental falls threaten the healthy life of people,especially the elderly living alone.In order to solve this problem,a human posture estimation system for home scenario is proposed.It mainly implements the functions of posture estimation,fall detection,monitoring and warning,and information management.The Center Net algorithm is selected to implement human posture estimation,and the convolution operator of the model is improved to address the situation that Center Net has missed and mis-detected nodes in home scenes,and pictures in home scenes are collected,and a training set with sufficient data volume and rich scenes is constructed through data annotation,label conversion and merging with COCO dataset.The improved algorithm can effectively improve the accuracy of the original Center Net model.Improve the application effect in home scenes.It is used for fall detection with better robustness.Using the pose estimation algorithm to obtain human skeleton information,the target is discriminated by calculating three pose features: human aspect ratio,height change rate,and torso tilt angle.Early warning information is sent when fall behavior is detected to provide timely help and rescue for the target.After completing the requirement analysis of the system,the overall architecture is constructed,each functional module of the system is divided and designed,and the human posture estimation system for home scenes is implemented using Py Qt5 framework.Test the function and performance of the system with good stability and scalability.The pose estimation system for home scenes has high accuracy and real-time test results in real home scenes.It has some practical value.更多还原