| 席志国,刘光辉.基于区域信息聚合的轻量化人群计数方法[J].高技术通讯(中文),2024,34(9):945~959 |
| 基于区域信息聚合的轻量化人群计数方法 |
| Lightweight crowd counting method based on regional information aggregation |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 09. 004 |
| 中文关键词: 人群计数； 区域信息聚合； 轻量化； 自注意力； 损失函数 |
| 英文关键词: crowd counting, regional information aggregation, lightweight, self-attention, loss function |
| 基金项目: |
| 作者 | 单位 | | 席志国 | （西安建筑科技大学信息与控制工程学院西安 710055）
（建筑机器人陕西省高等学校重点实验室西安 710055）
（西安市建筑制造智动化技术重点实验室西安 710055） | | 刘光辉 | |
|
| 摘要点击次数: 289 |
| 全文下载次数: 234 |
| 中文摘要: |
| 针对高密人群图像中细节信息丢失、背景噪声易与人群特征混淆以及网络模型复杂度高等问题，本文提出一种基于区域信息聚合的轻量化人群计数方法。首先，为获取高密图像中细粒度化的多尺度特征，设计了基于通道激活的多尺度特征提取模块，此模块通过引入Ghost卷积构建了层间分级类残差连接结构，同时对每级特征辅以通道激活，以轻量化的方式实现了网络感受野的逐级扩张。其次，提出一种自注意力区域信息聚合模块获取不同尺度区域的特征信息，该模块通过轻量级自注意力机制分别从通道和空间维度集成区域信息，增强对人群特征的关注，从而弱化背景噪声对计数的影响。最后，考虑到原始计数损失收敛过程中的不稳定性，在DM-Count损失的基础上引入一种新型计数损失，提高了模型稳定性和计数敏感性，进一步提升了计数性能。在Shanghai Tech、UCF-QNRF、JHU-CROWD++以及NWPU-Crowd这4个公开数据集的实验结果表明，本文所提方法对比其他主流轻量级人群计数方法有一定的提升，且模型参数量保持在较低水平。 |
| 英文摘要: |
| Aiming at the problems of loss of detailed information in high-density crowd images, background noise being easily confused with crowd features, and high complexity of network models, this article proposes a lightweight crowd counting method based on regional information aggregation. First, a multi-scale feature extraction module based on channel activation is designed to obtain fine-grained multi-scale features in high-density images. This module introduces Ghost convolution to construct an inter-layer hierarchical residual connection structure, and supplementing each level of features with channel activation, achieving a gradual expansion of the network’s receptive field in a lightweight manner. Secondly, a self-attention region information aggregation module is proposed to obtain feature information from regions of different scales. This module integrates region information from both channel and spatial dimensions using a lightweight self attention mechanism, enhancing focus on crowd features to weaken the impact of background noise on counting. Finally, considering the instability in the convergence process of the original count loss, a new counting loss is introduced based on the DM-Count loss, which improves the model stability and counting sensitivity, and further improves the counting performance. Experimental results on four public data sets of Shanghai Tech, UCF-QNRF, JHU-CROWD++, and NWPU-Crowd show that the method proposed in this thesis has a specific improvement compared with other mainstream lightweight crowd counting methods, and the number of model parameters remains relatively low-level. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
