Recently, a new IoT structure known as the Artificial Intelligence of Things (AIoT) comes into play. Crowd counting is a promising field in data analysis of AIoT, however, due to poor transparency and high data security risks, developing a novel network architecture that can precisely elevate the counting of heavy crowd is extremely difficult. In addition, the fusion of IoT and AI also poses several challenges. The focus of this work is on the effective design of IoT framework and deep learning algorithm towards security of smart city. The system can be used to estimate the crowd traffic in public places, and can prevent the occurrence of congestion, stampede and other accidents, such as stations, airports, large-scale exhibitions, tourist attractions and other places. The constructed system contains video collection, upload and display as well as data analysis and early warning operation at the embedded device end, and automatically tracks densely crowd areas by controlling the video monitoring device. Moreover, the cloud platform can be controlled through the network. Our proposed algorithms are composed of two main aspects, i.e., division and focus. Firstly, we propose a novel density-adaptive Gaussian kernel to elevate the quality of density maps. Then, we propose a module based on conditional random fields for feature fusion. Finally, we propose a block segmentation module to predict our segmentation results and extract the context-aware information in segmentation stage. Experiments on our captured data, the Shanghai Tech, UCF_CC_50 and UCF_QNRF datasets demonstrate that our solution has obtained better performance and lower count errors over the state of the art.

最近，一种被称为物联网 (AIoT) 的新物联网结构开始发挥作用。人群计数是AIoT数据分析的一个有前景的领域，但由于透明度差和数据安全风险高，开发一种能够精确提升大量人群计数的新型网络架构非常困难。此外，物联网和人工智能的融合也带来了一些挑战。这项工作的重点是针对智慧城市安全的物联网框架和深度学习算法的有效设计。该系统可用于公共场所的人流量估算，可预防车站、机场、大型展览、旅游景点等场所发生拥堵、踩踏等事故。构建的系统包含视频采集，在嵌入式设备端进行上传显示以及数据分析和预警操作，通过控制视频监控设备自动跟踪人群密集区域。而且，云平台可以通过网络进行控制。我们提出的算法由两个主要方面组成，即分割和聚焦。首先，我们提出了一种新的密度自适应高斯核来提高密度图的质量。然后，我们提出了一个基于条件随机场的模块进行特征融合。最后，我们提出了一个块分割模块来预测我们的分割结果并在分割阶段提取上下文感知信息。对我们捕获的数据进行实验，上海科技，