基于多层级语义信息融合的区域一致性半监督人群计数研究

Research on Regional Consistency Semi-Supervised Crowd Counting Based on Multi-Level Semantic Information Fusion

摘要: 面向智慧城市场景的人群计数问题，本文提出一种基于多层级语义特征提取网络的半监督人群计数框架。所提方法以ResNet-34为骨干网络，设计自上而下的多层级语义增强模块，在不同层级间通过门控进行语义信息增强与轻量级特征融合，抑制背景噪声、突出前景人群信息。在无标注样本上引入区域一致性约束，设计半监督框架，提升模型的泛化能力。训练阶段采用先监督训练预热，而后交替迭代更新的策略。在ShanghaiTech数据集和UCF_CC_50数据集上进行了训练和测试，实验结果显示，该框架在ShanghaiTech数据集A和B部分上的MAE分别为65.4和9.2，在UCF_CC_50数据集上的MAE为201.2，算法在不同密度人群与复杂背景场景下均表现出了较好的识别精度，可以为基于视觉的人群计数任务提供高效的解决方案。

Abstract: Aiming at the problem of crowd counting in smart city scenes, this paper proposes a semi-supervised crowd counting framework based on multi-level semantic feature extraction network. The proposed method uses ResNet-34 as the backbone network and designs a top-down multi-level semantic enhancement module. Semantic information enhancement and lightweight fusion are performed through gating between different levels to suppress background noise and highlight foreground crowd information. By introducing regional consistency constraints on unlabeled samples, a semi-supervised framework is designed to improve the model’s generalization ability. In the training stage, the strategy of supervised training preheating and then alternating iterative updating is adopted. The framework is trained and tested on the ShanghaiTech dataset and the UCF_CC_50 dataset. The experimental results show that the MAE of the framework on the A and B parts of the ShanghaiTech dataset is 65.4 and 9.2, respectively, and the MAE on the UCF_CC_50 dataset is 201.2. The algorithm shows good recognition accuracy in different crowd densities and complex background scenarios. This algorithm can provide an efficient solution for vision-based crowd counting tasks.