高校课堂密集遮挡场景下的动态多头目标感知

谢日敏; 谢大同; 申亮

文章摘要

高校课堂密集遮挡场景下的动态多头目标感知

Dynamic Multi-Head Object Perception in Heavily Occluded University Classrooms

投稿时间：2025-05-06 修订日期：2025-06-02

DOI：

中文关键词: 目标检测 YOLO FasterNet DyHead DyT

英文关键词: Object Detection YOLO FasterNet DyHead DyT

基金项目:福建省自然科学“高校理论课堂学生坐姿行为识别”(2021J01334).

作者	单位	邮编
谢日敏^*	福建商学院信息工程学院、通信与物联网工程系	350012
谢大同	福建商学院信息工程学院、计算机科学与技术系	350012
申亮	福建商学院信息工程学院、通信与物联网工程系	350012

摘要点击次数: 202

全文下载次数: 0

中文摘要:

针对高校课堂密集场景中存在的边缘目标特征模糊、遮挡导致漏检率高等问题,本文提出FasterDD网络框架。FasterDD首先利用FasterNet的?部分卷积(PConv)?和?动态拓扑结构,实现计算效率与表征能力的平衡,接着引入动态双曲正切函数(DyT) ,通过特征幅值自适应调节替代传统归一化,简化了计算过程,然后使用统一的动态目标检测头(DyHead)自适应融合多尺度特征,实现构建动态感受野。实验表明,FasterDD在自建的Classroom课堂数据集中,较YOLO11n(baseline) mAP50提升了 2.1%,mAP50-95 提升了 1.9%。推理速度保持190FPS(640×640输入),能够优化密集人群场景下的个体识别,提升关键区域的目标感知能力。

英文摘要:

To address the challenges of blurred edge features and high miss detection rates caused by occlusion in densely crowded university classroom scenes, this paper proposes the FasterDD network framework. FasterDD first leverages Partial Convolution (PConv) and the dynamic topology structure from FasterNet to balance computational efficiency and representational capacity. Then, a Dynamic Hyperbolic Tangent Function (DyT) is introduced to adaptively regulate feature amplitudes, replacing traditional normalization and simplifying the computational process. Finally, it incorporates a unified Dynamic Detection Head (DyHead) to adaptively fuse multi-scale features, enabling the construction of dynamic receptive fields for improved feature representation. Experiments on a self-constructed Classroom dataset show that FasterDD improves mAP50 by 2.1% and mAP50-95 by 1.9% over the YOLO11n baseline, while maintaining an inference speed of 190 FPS with 640×640 inputs. The proposed framework effectively enhances individual recognition in densely populated classroom scenarios and significantly improves target perception in critical regions.

View Fulltext 查看/发表评论下载PDF阅读器

关闭