近年来，平安城市的建设正在如火如荼地开展，但是在平安城市运转的过程中，难免会出现一些异常事件或不和谐的因素。为了精准快速地处理这些异常事件，保证平安城市的平稳运营，建立高效的异常检测与调度模型是重中之重。目前，人们广泛采用机器学习算法来构建视频异常检测模型。然而，大多数的异常检测模型存在计算复杂度高、需要大量计算资源等问题，因此边缘摄像头采集视频后需将视频回传至数据中台进行统一处理，这不仅对传输链路造成巨大压力，还显著提升了数据中台的计算和运行成本。针对检出的异常，数据中台需要调度工作人员进行异常处理。但是现有的调度算法考虑了过多的细节因素，这既不便于统计和管理，也提高了算法运行的复杂度。为此，本文提出了一种高效且轻量级的视频异常检测模型和异常事件处理模型，来准确、快速地检测异常，并高效地调度人员进行异常处理。本论文的研究贡献主要分为以下三个方面：（1）本文提出了一种基于受限玻尔兹曼机的轻量级视频异常检测模型。该模型利用FastFCN对视频帧进行目标检测，并将目标检测的结果进行切片作为受限玻尔兹曼机的输入。通过比较受限玻尔兹曼机输出结果的自由能量值和阈值的大小，模型就能判定该视频帧中是否存在异常。由于受限玻尔兹曼机仅有两层结构，因此，本模型计算复杂度低，能够直接被部署在边缘摄像头上，从而降低了链路传输和后端计算的压力。（2）本文设计了一种基于主成分分析法的异常事件处理模型，降低了调度算法运行的时间和开销，提高了异常处理的效率。通过筛选影响异常处理效率最为关键的6个影响因素，并邀请专家进行打分，得到针对所有因素的评分表。依据评分表，提取出主成分，从而实现数据的降维，并计算每个因素的综合得分。综合得分被用于计算每个因素的权值，从而得到最终的异常事件处理模型。（3）本文所提出的视频异常检测模型和异常事件处理模型已经在我国陕西省西安市幸福林带项目中进行了初步的系统验证和测试。测试结果表明，在真实的复杂条件下（如：天气、光线等），本文的视频异常检测模型能够稳定地进行异常检测；在处理异常事件的调度过程中，本文的异常事件处理模型能够满足在城域级复杂场景中进行高效异常处理的要求。
In recent years, the construction of peaceful cities is actively being carried out. However, during the operation of peaceful cities, there are inevitably some abnormal events or discordant factors that may occur. In order to accurately and quickly handle these abnormal events and ensure the smooth operation of peaceful cities, it is important to establish efficient abnormal detection and scheduling models. Currently, researchers widely use machine learning algorithms to build video abnormal detection models. However, most of the abnormal detection models have the problem of high computational complexity and need for a large amount of computational resources. In addition, after collecting video from edge cameras, the video needs to be sent back to the data center for unified processing. This not only creates a heavy burden on the transmission chain, but also significantly increases the computational and operating costs of the data center. In response to detected abnormalities, the data center needs to schedule personnel for abnormal processing. However, existing scheduling algorithms consider too many detailed factors, which is not convenient for statistics and management, and also has high complexity. Therefore, this thesis proposes an efficient and lightweight video abnormal detection model and abnormal event processing model to accurately and quickly detect abnormalities and efficiently schedule personnel for abnormal processing. The research contributions of this thesis mainly lie in the following three aspects:(1) This thesis proposes a lightweight video abnormal detection model based on a restricted Boltzmann machine (RBM). The model uses FastFCN to perform object detection on video frames and slices the detection results as the input of the RBM. By comparing the energy value of the RBM’s output result with a threshold, the model can determine whether there is an abnormality. Since the RBM has only two layers, the computational complexity of this model is low, and it can be directly deployed on edge cameras, thereby reducing the burden on the transmission chain and the backend computation.(2) This thesis designs an abnormal event processing model based on principle component analysis (PCA) method, which reduces the running time and overhead of the scheduling algorithm and improves the efficiency of abnormal processing. Through the selection of the six most crucial factors that affect the efficiency of abnormal processing and the invitation of experts to rate them, a rating table can be obtained for all factors. Based on the rating table, the main components can be extracted to achieve data dimensionality reduction and the calculation of the comprehensive score of each factor. The comprehensive score is used to calculate the weight of each factor and obtain the final abnormal event processing model.(3) The video abnormal detection model and abnormal event processing model proposed in this thesis have been partially verified and tested in the Xingfu Lindai Project in Xi‘an, China. The testing results show that the video abnormal detection model proposed in this thesis can stably detect abnormalities under real complex conditions such as weather (e.g., sunny and rainy days) and lighting (e.g., day and night). When processing abnormal events, the abnormal event processing model proposed in this thesis can meet the requirements of efficient abnormal processing in complex urban scenarios.
