Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (23): 280-286.DOI: 10.3778/j.issn.1002-8331.2105-0362

• Engineering and Applications • Previous Articles     Next Articles

Real-Time Construction Worker Detection Method for Edge Device

XU Xiaodong, WANG Junjie   

  1. School of Engineering, Ocean University of China, Qingdao, Shandong 266100, China
  • Online:2021-12-01 Published:2021-12-02

面向边缘端的施工人员实时检测方法

徐晓东,王俊杰   

  1. 中国海洋大学 工程学院,山东 青岛 266100

Abstract:

Construction worker detection has an important application value in construction management. The complex images backgrounds and diverse perspectives at construction sites make the task of construction worker detection difficult, while most of the construction sites have poor infrastructure facilities and poor network conditions, making them unsuitable for model deployment on large GPU workstations. To address the above issues, it proposes a real-time worker detection algorithm for construction scenes based on YOLOv3 detection network, adding a feature pyramid pooling module, increasing multi-scale feature fusion and improving candidate frames to improve detection accuracy, and using a channel pruning algorithm to lighten the detection network to adapt to the computing power of edge-end devices. This method achieves an average accuracy of 88.23% on the homemade construction worker dataset, which is 4.89 percentage points higher than the YOLOv3 detection method, and compresses the model size to 1/13 of the original one, doubles the detection speed. It achieves a detection speed of 69.08 frame/s on the embedded end device, meeting the requirements for real-time edge-end detection at construction sites.

Key words: pedestrian detection, YOLOv3, edge computing, channel pruning

摘要:

施工人员检测在施工管理工作中有重要的应用价值。施工现场图像背景复杂且视角多样,给施工人员检测任务带来难度,同时施工现场大多基础配套设施不完善,并且网络条件较差,不适合在大型GPU工作站上进行模型部署。针对以上问题,以YOLOv3检测网络为基础,加入特征金字塔池化模块,增加多尺度特征融合并改进候选框,提升检测精度,同时采用通道剪枝算法对检测网络进行轻量化处理以适应边缘端设备算力,提出一种面向边缘端的施工人员实时检测方法。该方法在自制的施工人员数据集上平均准确率可达到88.23%,较YOLOv3检测方法提升4.89个百分点,且将模型大小压缩至原来的1/13,检测速度提升一倍,在嵌入式设备Jetson Xavier NX上检测速度可达到69.08帧/s,满足在施工现场进行实时边缘端检测的要求。

关键词: 人员检测, YOLOv3, 边缘计算, 通道剪枝