用于课堂行为识别的轻量化姿态估计网络研究  

Research on lightweight pose estimation networks for classroom behavior recognition

作　　者：薛涛[1] 曹哲睿 周千明[1] XUE Tao;CAO Zherui;ZHOU Qianming(School of Computer Science,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学计算机科学学院,陕西西安710048

出　　处：《西安工程大学学报》2024年第6期124-134,共11页Journal of Xi’an Polytechnic University

基　　金：陕西省教育厅科研项目基金(19JK0377)。

摘　　要：课堂行为识别需要模型具有较快的推理速度、较少的参数量以及较好的模型鲁棒性。现有轻量化模型虽然达到了不错的实时性能,但其在面对课堂复杂场景时准确率较低,难以达到实际要求。针对此问题,提出了聚焦线性注意力姿态估计网络FLAPose(focused linear attention pose,FLAPose),重新设计的分组注意力(sectionalization attention,SA)增强了模型捕捉局部信息的能力,然后使用聚焦线性注意力(focused linear attention,FLA)设计增强模型学习骨架信息的效果,最后通过骨骼损失对模型辅助监督,增强模型对遮挡重叠区域的学习能力。实验表明:与基线网络RTMPose相比,FLAPose在更少参数和计算量的情况下,COCO数据集上的平均准确率提升了1.7%,CS-Dataset数据集上平均准确率提升4.8%。此外,模型通过TensorRT加速后部署在推理服务器中,测量了模型在GPU上的每秒帧数(frames per second,FPS),FLAPose的FPS在GPU上超过764.460,在CPU上超过215.63,高于其他所有网络。Classroom behavior recognition requires models to have fast inference speed,fewer parameters,and good robustness.Although existing lightweight models achieve commendable real-time performance,they fall short in accuracy when dealing with the complex scenarios typical of classroom environments,failing to meet practical requirements.To address this issue,we proposed the focused linear attention pose estimation network(FLAPose),which redesigned sectionalization attention(SA)to enhance the model′s capability to capture local information.Subsequently,focused linear attention(FLA)was employed to improve the model′s ability to learn skeletal information.Finally,bone loss was introduced as an auxiliary supervision mechanism to strengthen the model′s learning capacity in occluded and overlapping regions.The experimental results show that FLAPose outperforms the baseline network RTMPose by achieving an average accuracy improvement of 1.7%on the COCO dataset and 4.8%on the CS-Dataset,a model with fewer parameters and computational resources.Moreover,by leveraging TensorRT for acceleration and deploying the model on an inference server,we measured the frames per second(FPS)of the model on both GPU and CPU.FLAPose achieves an FPS exceeding 764.460 on the GPU and over 215.63 on the CPU,surpassing all other networks.

关 键 词：课堂行为识别 轻量级 线性注意力 姿态估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]