基于多维度感知特征金字塔的人群计数算法  

A Multi-Dimensional Perception Feature Pyramid-Based Algorithm for Crowd Counting

作　　者：陈慧[1] CHEN Hui(School of Integrated Circuits,Wuxi Institute of Technology,Wuxi 214121,China)

机构地区：[1]无锡职业技术学院集成电路学院,江苏无锡214121

出　　处：《南通职业大学学报》2025年第1期72-79,共8页Journal of Nantong Vocational University

摘　　要：针对人群计数中存在的尺度多变和人头误判问题,提出一种基于多维度感知特征金字塔的人群计数算法。该算法以特征金字塔编解码网络为基础,由特征聚合模块组成高效双解码结构,通过多次融合相邻层次语义信息,保留不同尺度下的细节特征,更好地适应了人头的尺度变化。此外,在网络视野最高处引入多维度感知模块,从空间和通道等多个维度汇聚人头关键特征,更新不同位置下的特征权重,将人头信息从背景中有效加以区分,进一步缩小了单个目标的预测范围。采用多层次监督进行网络整体训练,定性与定量分析结果表明,所提算法在四个公共数据集上表现达到预期。To address the challenges of scale variation and head misidentification in crowd counting,a multi-dimensional perception feature pyramid-based crowd counting algorithm is proposed.Building upon a feature pyramid encoder-decoder network,the algorithm uses feature aggregation modules to construct an efficient dual-decoding structure.By repeatedly fusing semantic information from adjacent hierarchical levels,the network retains fine-grained features across different scales to better adapt to head size variations.Furthermore,a multi-dimensional perception module is incorporated at the apex of the network,which aggregates key head features from multiple dimensions,such as spatial and channel,and updates the feature weights at different positions.This not only effectively distinguishes head information from background but also further narrows prediction ranges for individual targets with multi-level supervision for holistic network training.The results of qualitative and quantitative analyses show that the proposed algorithm achieves superior performance across four public benchmark datasets.

关 键 词：人群计数 多维度感知特征金字塔 卷积神经网络 注意力机制 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]