Crowd Gathering Detection Method Based on Multi-Scale Feature Fusion and Convolutional Attention

Kamil Yasen, Jingyang Zhou, Nan Zhou, Ke Qin, Zhiguo Wang, Ye LiSensors 2025 · DOI

TL;DR

This paper introduces MSCANet, a novel deep learning framework for crowd gathering detection that tackles challenges like occlusion and varying crowd densities. By integrating multi-scale feature fusion and a convolutional attention mechanism, MSCANet significantly improves detection accuracy and robustness in complex scenarios, making it a promising tool for public safety.

What Problem Does This Paper Address?

As urban populations grow, so does the frequency and density of crowd gatherings, posing significant challenges for public safety management. Existing crowd detection methods often struggle with:

  • Limited Contextual Understanding: Many methods rely solely on local texture or density features, failing to capture broader scene context, which is crucial for accurate detection, especially in occluded or complex backgrounds.
  • Fixed-Scale Limitations: Standard feature extraction techniques can't effectively adapt to the wide range of crowd densities and scales present in real-world scenes.
  • Difficulty with Dense Crowds: Densely populated areas are particularly challenging, as it's hard to distinguish individual heads or capture critical local features.
  • Lack of Generalization: A universal point-supervised model that performs well across different datasets and scenarios remains elusive.

This paper aims to address these limitations by developing a more robust and adaptable crowd gathering detection system.

Methods

The proposed solution is a point-supervised framework called MSCANet (Multi-Scale Convolutional Attention Network). Its core components include:

  • Multi-Scale Feature Extraction and Fusion: MSCANet builds upon the VGG16 backbone and a feature pyramid module (similar to P2PNet) to extract multi-scale feature maps. These are then fused to create a richer representation.
  • Attention Enhancement Module: This is a key innovation. It incorporates both channel attention (focusing on important feature channels) and spatial attention (highlighting critical spatial locations). This module helps the network suppress irrelevant features and enhance salient ones, improving representation, especially in occluded scenes.
  • Attention Guided Scale Adaptation (AGSA) Module: This module further refines the attention mechanism by using four parallel channels, each with dilated convolutions of different rates (1-4) followed by the attention enhancement module. This allows for more effective capture of multi-scale contextual information and better handling of scale variations.
  • Point-Supervised Localization and Regression: Similar to prior work, MSCANet uses point annotations. It predicts candidate points and their confidence scores. A one-to-one matching mechanism (using the Hungarian algorithm) then selects the final predicted points based on distance and confidence.

Datasets

The paper primarily uses the following datasets for training and evaluation:

  • ShanghaiTech PartA: Known for its complex scenes, uneven crowd distributions, and large scale variations.
  • ShanghaiTech PartB: Characterized by lower crowd density, more uniform scales, and simpler backgrounds, often featuring urban street scenes.

The authors also conducted experiments on a merged dataset combining PartA and PartB to simulate more complex real-world scenarios and evaluate generalization.

Key Insights and Ideas

  • Attention is Key for Occlusion: The integration of channel and spatial attention mechanisms proved crucial for improving feature representation and handling occluded targets. Simply removing either attention component led to a performance drop.
  • Multi-Scale Fusion is Essential: Combining features from different scales allows the model to adapt to varying crowd densities, from sparse to extremely dense.
  • Point Supervision Can Be Powerful: By directly predicting head locations (via offsets from reference points) and confidence scores, the model bypasses the need for bounding boxes, which can be error-prone with weak supervision.
  • Generalization is Achievable: The proposed model demonstrates strong generalization capabilities, performing well even when trained on one dataset and tested on another, or when trained on a fused dataset.

Key Results

  • Superior Performance on ShanghaiTech: MSCANet achieved the best performance on ShanghaiTech PartA, with the lowest MAE and MSE. It also showed competitive results on PartB, achieving the best MAE and second-lowest MSE.
  • Significant Improvements over P2PNet: Compared to P2PNet (a leading point-supervised method), MSCANet reduced MAE by 2% and MSE by 3.9% on PartA.
  • Robustness in Cross-Validation: In cross-validation experiments, MSCANet consistently outperformed P2PNet and other baseline methods, demonstrating its ability to adapt to varied scene complexities and crowd densities. For instance, in cross-validation between PartA and PartB, MSCANet's MAE and MSE were 3.9 and 5.3 lower than P2PNet, respectively.
  • Ablation Studies Confirm Module Effectiveness: The ablation studies confirmed that using three AGSA modules in both branches and combining channel and spatial attention yielded the best results.

Limitations and Future Work

  • Computational Cost: While effective, the attention mechanisms and multi-scale processing can increase computational complexity. The paper doesn't explicitly detail inference speed.
  • Hyperparameter Tuning: The performance is sensitive to hyperparameters like the learning rate and the balance weight for loss functions, requiring careful tuning.
  • Future Work: The authors suggest that the model's effectiveness in complex scenes and its strong generalization ability make it suitable for a wider range of applications beyond just crowd counting, such as group aggregation event detection and general scene monitoring. Further exploration into real-time performance and more diverse datasets could be future directions.

Summary updated

Abstract

With rapid urbanization and growing population inflows into metropolitan areas, crowd gatherings have become increasingly frequent and dense, posing significant challenges to public safety management. Although existing crowd gathering detection methods have achieved notable progress, they still face major limitations: most rely heavily on local texture or density features and lack the capacity to model contextual information, making them ineffective under severe occlusions and complex backgrounds. Additionally, fixed-scale feature extraction strategies struggle to adapt to crowd regions with varying densities and scales, and insufficient attention to densely populated areas hinders the capture of critical local features. To overcome these challenges, we propose a point-supervised framework named Multi-Scale Convolutional Attention Network (MSCANet). MSCANet adopts a context-aware architecture and integrates multi-scale feature extraction modules and convolutional attention mechanisms, enabling it to dynamically adapt to varying crowd densities while focusing on key regions. This enhances feature representation in complex scenes and improves detection performance. Extensive experiments on public datasets demonstrate that MSCANet achieves high counting accuracy and robustness, particularly in dense and occluded environments, showing strong potential for real-world deployment.
BibTeX
@inproceedings{yasen2025crowd,
  author = {Kamil Yasen and Jingyang Zhou and Nan Zhou and Ke Qin and Zhiguo Wang and Ye Li},
  title = {{Crowd Gathering Detection Method Based on Multi-Scale Feature Fusion and Convolutional Attention}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25216550},
}
Figure