【作者】 李超；

【导师】 李振华；

【作者基本信息】 山东大学 ， 控制工程（专业学位）， 2023， 硕士

【摘要】 随着人工智能技术的发展和消费者日益增长的智能化家居需求,将深度学习技术与智能化设备的控制调节相结合,能够有效地推动智能家居领域向着主动智能化方向发展。同时,人们也越来越追求健康的生活方式,运动、健身已成为现代人必不可少的一部分,对于运动量监测及评价的需求日益增加。基于以上需求,本文以室内家居环境的人体目标检测和人体运动量分析为任务背景,研究了深度学习模型在嵌入式平台的应用实现。本文的主要工作如下:(1)针对深度卷积神经网络模型参数计算量大、训练时间长、模型权重大等问题,构建了一种新的轻量化网络模型MobileOne-YOLOv5-lite改进网络进行人体目标检测,实现了较好的目标检测效果。对比实验表明MobileOne-YOLOv5-lite模型的计算量与参数量均大幅度减少,GPU硬件环境下检测准确率高达98.0%,单位样本预测时间仅为9ms。(2)本文提出了一种新的基于摄像头的地面网格公式单目测距方法,结合第一阶段的人体目标检测结果,根据该身体框对应的网格测距公式来计算距离,实验结果表明在4m 4m的区域内测距平均误差为34.96cm,为下一阶段人体运动量分析奠定了基础。(3)针对AlphaPose算法关键点提取模型计算量和参数量规模庞大问题,本文设计了轻量化Mobile-AlphaPose网络;同时针对非规则行为的人体运动量分析需求,设计了人体运动量估计的双重判断系统,将人体运动行为分为规范标准型和非规范标准型,分情况分析不同动作行为的运动量大小。对于规范标准型行为采用MLP行为识别分类网络对关键点姿态预测出所属的行为类别并得出其运动量级别,针对非规范标准行为的运动量估计提出了依据人体移动距离快慢和肢体摆动幅度快慢衡量运动量大小的方案。经过实验验证表明,轻量版的Mobile-AlphaPose网络在嵌入式平台下检测帧率达到了17.8frame/s,行为识别分类准确率达到了 85.7%,基本满足检测和识别需求。(4)在上述研究的基础上,完成了算法模型在嵌入式端的匹配应用。以瑞芯微RV1126为处理器完成了嵌入式平台的系统固件制作和开发环境搭建,并将深度学习模型部署到嵌入式端进行应用。通过不同硬件环境下模型运行效果对比,实现了在嵌入式平台检测精度和效率的平衡,验证了深度学习模型在嵌入式平台应用的有效性,体现其实用价值。更多还原

【Abstract】 With the development of artificial intelligence technology and the increasing demand of consumers for smart home,the combination of deep learning technology and the control and adjustment of intelligent equipment can effectively promote the development of smart home towards the direction of active intelligence.At the same time,people are increasingly pursuing a healthy lifestyle.Exercise and fitness have become an indispensable part of modern people,and the demand for monitoring and evaluation of exercise amount is increasing.Based on the above requirements,this paper studies the application and realization of deep learning model on embedded platform with the background of human target detection and human movement analysis in indoor home environment.The main work of this paper is as follows:(1)Aiming at the problems such as large calculation amount of parameters,long training time and heavy model weight of deep convolutional neural network,a new lightweight network model MobileOne-YOLOv5-lite is constructed to improve the network for human target detection,and good target detection effect is achieved.The comparison experiment shows that the calculation amount and parameter number of MobileOne-YOLOv5-lite model are greatly reduced,the detection accuracy is up to 98.0%in GPU hardware environment,and the prediction time per unit sample is 9ms.(2)This paper proposes a new ground grid formula monocular distance measurement method based on camera.Combined with the human target detection results in the first stage,the distance is calculated according to the grid distance measurement formula corresponding to the body frame.The experimental results show that the average distance measurement error is 34.96cm in the 4m×4m area,which lays the foundation for the next stage of human exercise analysis.(3)Aiming at the problem of large-scale calculation and parameter volume of the key point extraction model of the AlphaPose algorithm,this paper designs a lightweight MobileAlphaPose network;At the same time,a dual judgment system for human body movement estimation is designed to meet the needs of human body movement analysis in irregular behaviors.Human movement behavior is divided into normative standard type and nonstandard standard type,and the amount of exercise of different action behaviors is analyzed according to the situation.For the standard behavior,MLP behavior recognition classification network is used to predict the behavior category of the key point posture and get its level of exercise.For the estimation of the amount of exercise of the non-standard behavior,a scheme is proposed to measure the amount of exercise according to the speed of human moving and the speed of the limb swing.Experimental verification shows that the detection frame rate of the lightweight version of the Mobile-AlphaPose network on the embedded platform has reached 17.8 frame/s,and the accuracy rate of behavior recognition and classification has reached 85.7%,basically meeting the detection and recognition requirements.(4)On the basis of the above research,the matching application of the algorithm model on the embedded side is completed.With Rockchip RV1126 as the processor,the system firmware production and development environment of the embedded platform are completed,and the deep learning models are transplanted to the embedded terminal for application.Through the comparison of the model running effects under different hardware environments,the balance between detection accuracy and efficiency on the embedded platform is achieved,and the effectiveness of the deep learning model in the embedded platform application is verified,reflecting its practical value.更多还原