摘要
针对复杂场景下智能监控系统存在的遮挡敏感和多目标检测精度不足问题,构建了多传感器融合的嵌入式监控系统。在硬件层面,搭建了基于系统级芯片的嵌入式架构;在算法层面,提出增强型YOLOv8n模型,引入高效局部注意力模块混合结构特征金字塔网络(Eficient Local Attention-Hybrid Structure Feature Pyramid Network,ELA-HSFPN)来优化多尺度特征融合,引入聚焦交并比损失函数来提升边界框回归精度。改进后模型参数量、计算量和体积均降低了15.2%,平均精度均值(mean Average Precision,mAP)提升了3%。实验结果表明,系统在0.17 s内可完成目标检测,跌倒识别、目标分类准确率超过95%,可有效检测70 cm范围内遮挡目标。该系统在计算效率、检测精度和抗遮挡性能方面均优于传统方案,为复杂室内环境下的实时安全监控提供了可靠的解决方案。
To address the issues of occlusion sensitivity and insufficient multi-object detection accuracy in intelligent monitoring systems for complex scenarios,an embedded monitoring system based on multi-sensor fusion is presented.At the hardware level,an embedded architecture based on system-on-chip is constructed.At the algorithm level,an enhanced YOLOv8n model is proposed,an efficient local attention mixed-scale feature pyramid module is introduced to optimize multi-scale feature fusion,and a focused intersection-over-union loss function is applied to improve bounding box regression accuracy.The improved model reduces parameters,computational load,and model size by 15.2%,while increasing the mean average precision(mAP)by 3%.Experimental results show that the system completes object detection within 0.17 s,achieves fall recognition and object classification accuracy exceeding 95%,and effectively detects occluded targets within a 70 cm range.The proposed system outperforms traditional solutions in computational efficiency,detection accuracy,and anti-occlusion performance,providing a reliable real-time safety monitoring solution for complex indoor environments.
作者
潘家亮
李博
阮斌
PAN Jialiang;LI Bo;RUAN Bin(School of Physics,Zhejiang University of Technology,Hangzhou 310023,China;Zhejiang Uniview Technologies Co.,Ltd.,Hangzhou 310051,China)
出处
《测控技术》
2025年第9期52-60,共9页
Measurement & Control Technology
基金
浙江工业大学产学研项目(KYY-HX-20220057)。
