唐其伟 杜卉然 程伟 陈刚
（日立楼宇技术（广州）有限公司，广东 广州 510700）
摘要：针对电梯轿厢场景下，基于传统RGB图像实现人体目标检测易受电梯轿厢内饰、光照、乘客外貌特征等因素影响，难以实现稳定、准确的乘客数量检测的问题，提出一种基于深度图像的人体目标检测方法。首先，设计深度图像采集装置，构建人体目标深度图像数据集；然后，提出一种适配专用推理架构的限制采样尺度及线性激活的YOLOv8网络优化方法。经实验验证，该方法在边缘计算平台上具有0.984的AP@0.5准确率及44 ms的推理延时，满足电梯轿厢场景下人体目标实时检测的需求，提升了电梯的运行效率及安全性。
关键词：电梯轿厢场景；深度图像；人体目标检测；边缘计算
中图分类号：TP391.41 文献标志码：A 文章编号：1674-2605(2024)06-0011-08
DOI：10.3969/j.issn.1674-2605.2024.06.011 开放获取
Human Object Detection Method Based on Depth Images
TANG Qiwei DU Huiran CHENG Wei CHEN Gang
(Hitachi Building Technology (Guangzhou) Co., Ltd., Guangzhou 510700, China)
Abstract: Aiming at the problem that human object detection based on traditional RGB images in elevator car scenes is easily affected by factors such as elevator car interior, lighting, and passenger appearance characteristics, making it difficult to achieve stable and accurate passenger quantity detection, a human object detection method based on depth images is proposed. Firstly, design a depth image acquisition device and construct a dataset of human target depth images; Then, a YOLOv8 network optimization method with limited sampling scale and linear activation adapted to a dedicated inference architecture is proposed. Experimental results show that this method has 0.984 AP@0.5 The accuracy and 44 ms inference delay meet the real-time detection requirements of human targets in elevator car scenes, improving the operational efficiency and safety of elevators.
Keywords: elevator car scene; depth image; human object detection; edge computing
