基于多任务学习的智慧城市场景人群计数研究

Multi-Task Learning-Based Crowd Counting for Smart-City Scenes

摘要: 为解决城市场景摄像头高度与视角差异导致的人群图像中强透视、多尺度问题，提出一种以多任务学习为框架的多尺度人群计数算法。在点监督框架上引入一种一体化多尺度金字塔模块，提升对不同大小人头的特征提取能力，同时将点标注经多尺度高斯扩散与自适应阈值自动生成人群前景分割标签，作为辅助任务与计数任务联合训练，以计数损失和分割损失构成多任务目标，抑制背景干扰并实现多尺度人群的识别。选取以城市场景为代表的ShanghaiTech数据集A、B和UCF-QNRF数据集进行实验，MAE分别为57.8，7.6和86.2，在强透视和密度不均场景下均取得较好的效果，表现出较高鲁棒性。所提方法以零额外标注成本显著提升了智慧城市场景下人群计数的准确性与可部署性，适用于跨视角的城市监控设备。

Abstract: To address the strong perspective distortion and multi-scale variations in crowd images caused by differences in camera height and viewpoint in urban settings, we propose a multi-scale crowd counting method under a multi-task learning (MTL) framework. On top of a point-supervised paradigm, we introduce an integrated multi-scale pyramid module that enhances feature extraction for heads of different sizes. Meanwhile, point annotations are converted—via multi-scale Gaussian diffusion and adaptive thresholding—into crowd-foreground segmentation labels, which serve as an auxiliary task jointly trained with the counting task. The overall objective combines counting loss and segmentation loss, suppressing background interference and enabling reliable recognition across scales. Experiments on urban-scene representative datasets—ShanghaiTech Part A, ShanghaiTech Part B, and UCF-QNRF—yield MAEs of 57.8, 7.6, and 86.2, respectively, demonstrating strong performance under severe perspective and uneven density conditions and indicating high robustness. The proposed approach improves accuracy and deployability for smart-city crowd counting without any additional annotation cost, making it suitable for cross-view urban surveillance systems.