【作者】 李帅；

【导师】 伍洲； 齐宏拓；

【作者基本信息】 重庆大学 ， 工程（控制工程）（专业学位）， 2020， 硕士

【摘要】 建筑行业发展方式比较粗放,建筑工人技能素质不高,监管体系落后,难以统计工人施工状态。在施工现场,识别工人施工状态,获取工人施工效率,可使管理人员精确地量化和基准化劳动生产率,从而保障施工的进度。当前建筑工地工人生产率评估完全依赖于人工观察,在复杂的施工现场获取工人的施工状态数据需要花费大量的时间和人力,并且无法实时统计出每个工人的施工情况。针对上述的问题,现在研究人员使用动作识别的算法来识别工人的活动,进行施工效率分析,但目前只能对单个工人进行分析,当施工人员数量增多时无法进行准确的分析,不能满足实际的需求。本文基于改进的姿势估计算法获取人体关键点,研究基于人体关键点数据进行工人的跟踪和动作的识别,获取工人的施工效率。制作了施工动作的关键点数据集,用来训练动作识别网络,搭建了识别系统,最后进行整体的测试。论文主要完成的工作如下:(1)本文采用人体关键点数据进行动作分析,考虑实际施工现场需要同时处理多路摄像头视频信息。对用于提取人体关键点数据的姿势估计网络Open Pose计算量大、处理速度慢的问题,进行轻量化改进。用Mobile Net V2取代VGG-19主干网络,可分离式卷积取代两个分支网络中的普通卷积,将改进网络在COCO数据集进行性能测试,准确度较原来有所下降,但在实际视频测试提取到关键点信息,结果差距不大,且实际视频处理速度提升。(2)考虑到工地摄像头抖动和工人遮挡问题,完成关键点跟踪算法设计,选择采用多人跟踪算法Deep Sort,利用关键点信息提取工人的边界框,完成运动信息和外观信息提取,有效的跟踪工人以及关键点信息。测试结果表明该算法适合工地场景的应用。(3)多角度采集工人施工动作,建立关键点动作数据集,从数据集中提取重心特征、角度特征以及具有时间信息的速度特征。考虑动作的时间连续性,对关键点进行时间序列分析,使用多层MLP网络及堆叠LSTM网络对关键点数据集进行训练。测试结果表明具有时间信息的数据集和堆叠LSTM网络的采用,提高了动作识别的准确度。同时对于系统进行了整体的测试,运行速度基本可以满足实际需要。更多还原

【Abstract】 The development mode of the construction industry is relatively extensive,the skills and quality of construction workers are not high,and the supervision system is backward,making it difficult to count the construction status of workers.At the construction site,identifying the workers’ construction status and obtaining the workers’ construction efficiency can enable managers to accurately quantify and benchmark labor productivity,thereby ensuring the progress of the construction.The current evaluation of workers’ productivity on construction sites relies entirely on manual observation.It takes a lot of time and manpower to obtain workers’ construction status data on complex construction sites,and it is impossible to calculate the construction status of each worker in real time.In response to the above problems,researchers now use motion recognition algorithms to identify workers’ activities and analyze construction efficiency.However,at present,only single workers can be analyzed.When the number of construction workers increases,accurate analysis cannot be performed and it cannot meet the actual demand.In this paper,based on the improved pose estimation algorithm to obtain the keypoint of the human body,the study is based on the data of the keypoint of the human body to track and identify the movement of the workers to obtain the construction efficiency of the workers.A keypoint dataset of construction actions was produced to train the action recognition network,a recognition system was built,and finally the overall test was conducted.The main work of the thesis is as follows:(1)This article uses human body keypoint data for motion analysis,considering that the actual construction site needs to process multi-channel camera video information at the same time.Lightweight improvements were made to the problem of large computational volume and slow processing speed of the pose estimation network Open Pose for extracting keypoint data of the human body.Replace the VGG-19 backbone network with Mobile Net V2,and the separable convolution replaces the ordinary convolution in the two branch networks.The improved network will be tested on the COCO data set.The accuracy is lower than the original,but it is extracted in the actual video test.To the keypoint information,the result is not big,and the actual video processing speed is improved.(2)Considering the problem of camera shake and worker occlusion on the construction site,we completed the design of the keypoint tracking algorithm,and chose to use the multi-person tracking algorithm Deep Sort,using the keypoint information to extract the worker’s bounding box,complete the extraction of motion information and appearance information,and effectively track the workers and keypoint Information.The test results show that the algorithm is suitable for the application on site.(3)Collect workers’ construction actions from multiple angles,establish keypoint action data sets,and extract center of gravity features,angle features,and speed features with time information from the datasets.Considering the time continuity of actions,time series analysis is performed on keypoint,and a multi-layer MLP network and a stacked LSTM network are used to train keypoint datasets.The test results show that the adoption of datasets with time information and stacked LSTM networks improves the accuracy of action recognition.At the same time,the system has been tested overall,and the running speed can basically meet the actual needs.更多还原