| 引用本文: | 林园园1,2,杨会成1,2,胡耀聪1,2.基于轻量化卷积神经网络的人数估计算法研究(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2026，43（1）：28-38 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| :目的 目前人群计数模型中存在两种问题:复杂的重型计数模型虽然计数性能较强,但模型参数量和计算量
过大,因此实用性不高;当前的轻量化模型虽然降低了模型的复杂度,但计数性能不佳。 针对以上问题,提出一种
有效均衡计数性能和计数效率的基于轻量化卷积神经网络的人群计数模型。 方法 该方法分为两个模块:特征提取
模块和密度图回归模块。 首先,在特征提取模块打破以往提取特征时丢弃高度相似信息的思想,更加注重本征特
征和相似特征的融合,设计了一个轻量化线性映射单元,在减少网络参数和计算成本的同时,提高了计数精度;然
后,由多个线性映射单元组成轻量化线性映射块,并串行多个线性映射块组成特征提取模块;接着,将特征提取模
块提取到的特征馈送到密度图回归模块,密度图回归模块不再使用较少的标准卷积来回归密度图,而是使用扩张
卷积来替代标准卷积,利用堆叠的扩张卷积来增加感受野从而得到更加精确的回归密度图;最后将回归密度图求
和得到估计人数。 结果 所提方法的参数量仅有 0. 12 MB(Mbyte),计算量仅有 9. 23 GFLOPS(Giga Floating-point
Operations per Second),与其余轻量化人群计数模型相比均有降低,且在 3 个人群计数数据集,即 Shanghai Tech 数
据集、UCF-QNRF 数据集、NWPU-Crowd 数据集都取得了较为优异的计数性能。 结论 模型在保证计数性能的同时
也保证了计数效率,实现了两者的最佳平衡,并实现了实时快速精确的人群计数,相较于其他轻量级人群计数算
法,拥有更高的计数性能和计数效率,更具备实用性。 |
| 关键词: 永磁同步电机 新型趋近律 快速非奇异终端滑模控制 扰动观测器 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Research on Crowd Counting Algorithm Based on Lightweight Convolutional Neural Network |
|
LIN Yuanyuan1 2，YANG Huicheng1 2， HU Yaocong1
|
|
1. School of Electrical Engineering Anhui Polytechnic University Wuhu 241000 Anhui China
2. Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment Ministry of Education Wuhu
241000 Anhui China
|
| Abstract: |
| Objective There are two problems in current crowd counting models. Complex heavy-weight counting models
have strong counting performance but their large number of model parameters and high computational complexity result in
low practicality. Current lightweight models reduce the model complexity but have poor counting performance. To address
these issues a crowd counting model based on a lightweight convolutional neural network is proposed to effectively balance
counting performance and counting efficiency. Methods The method consisted of two modules the feature extraction module and the density map regression module. First in the feature extraction module instead of discarding highly
similar information during feature extraction as in the past more attention was paid to the fusion of intrinsic features and
similar features. A lightweight linear mapping unit was designed which improved the counting accuracy while reducing
network parameters and computational costs. Then multiple linear mapping units formed a lightweight linear mapping
block and multiple linear mapping blocks were serially connected to form the feature extraction module. Next the
features extracted by the feature extraction module were fed into the density map regression module. Instead of using a
small number of standard convolutions to regress the density map the density map regression module used dilated
convolutions to replace standard convolutions. Stacked dilated convolutions were utilized to increase the receptive field
and a more accurate regression density map was obtained. Finally the estimated number of people was obtained by
summing the regressed density map. Results The proposed method had a model size of only 0. 12 MB and a computational
cost of only 9. 23 GFLOPS Giga Floating-point Operations per Second both of which were lower than those of other
lightweight crowd counting models. It also achieved excellent counting performance on three crowd counting datasets the
Shanghai Tech dataset the UCF-QNRF dataset and the NWPU-Crowd dataset. Conclusion The model ensures both
counting performance and counting efficiency achieving the best balance between the two. It realizes real-time fast and
accurate crowd counting. Compared with other lightweight crowd counting algorithms it has higher counting performance
and efficiency and is more practical. |
| Key words: permanent magnet synchronous motor novel reaching law fast non-singular terminal sliding mode control
disturbance observer |
