当今视频监控系统在经济发展、公共安全管控、构建和谐社会等方面发挥着越来越重要的作用，如何从海量的监控图像数据中及时准确地检测识别异常行为对提升管控效率和保障公共安全有着重要意义，同时也具有较大的技术难度。尽管计算力提升带来了新的发展机遇，视频监控系统面临的识别成本和行为识别难题仍旧存在，影响了人体异常行为的检测与识别效率。针对这些挑战，本文通过深入研究深度学习技术在人体异常行为检测与识别中的应用，提出了基于改进BoTNet网络和PoseC3D网络的算法。本文主要研究工作如下：
1. 针对人体姿态估计算法提取的骨架关键点信息与其他3D网络提取的RGB信息难以融合的问题，基于SlowFast思想提出一种改进PoseC3D网络的人体异常行为检测算法。首先，采用HRNet模型对人体异常行为视频进行目标检测并提取2D关键点；其次，利用PoseC3D模型将2D关键点堆叠为3D热图，并在ResNet3D网络中引入CBAM注意力模块，为模型提供时间和空间信息的有效捕获；最后，为了优化模型的整体性能，使用轻量高效的X3D架构作为检测头。实验结果表明，在HMDB51数据集上改进后的模型准确率达到68.58%。
2．针对传统异常行为检测中2D卷积网络无法捕捉时间维度关系以及3D卷积网络训练成本高昂的问题，提出一种基于改进Bottleneck Transformer（BoTNet）网络的人体异常行为检测算法。首先，在BoTNet骨干网络上添加时间位移模块（TSM）提取时空特征信息；其次，引入SimAM模块处理空间和时间维度特征；最后，采用softmax进行分类。实验结果表明，在HMDB51数据集和UCF101数据集中准确率分别达到 74.10%和94.40%。
3. 利用IntelliJ IDEA等开发平台，本系统采用Java语言设计并实现了人体异常行为检测系统。通过对实际数据的测试验证，本系统证明了其在公共场所监控系统中，作为辅助监测工具的有效性。
 

Nowadays, video surveillance systems play an increasingly important role in economic development, public safety control, and building a harmonious society. It is of great significance to timely and accurately detect and identify abnormal behaviors from massive surveillance image data to improve control efficiency and ensure public safety, while also having significant technical difficulties. Despite the new development opportunities brought by the improvement of computing power, video surveillance systems still face recognition costs and behavioral recognition challenges, which affect the efficiency of detecting and recognizing abnormal human behavior. In response to these challenges, this article proposes an algorithm based on improved BoTNet network and PoseC3D network through in-depth research on the application of deep learning technology in human abnormal behavior detection and recognition. The main research work of this article is as follows:
1. To address the issue of difficulty in integrating skeleton key point information extracted by human pose estimation algorithms with RGB information extracted by other 3D networks, an improved PoseC3D network based human abnormal behavior detection algorithm is proposed based on the SlowFast idea. Firstly, the HRNet model is used to detect human abnormal behavior videos and extract 2D key points; Secondly, the PoseC3D model is used to stack 2D key points into 3D heatmaps, and the CBAM attention module is introduced into the ResNet3D network to provide effective capture of temporal and spatial information for the model; Finally, in order to optimize the overall performance of the model, a lightweight and efficient X3D architecture is used as the detection head. The experimental results show that the improved model accuracy on the HMDB51 dataset reaches 68.58%.
2. A human abnormal behavior detection algorithm based on an improved Bottleneck Transformer (BoTNet) network is proposed to address the issues of 2D convolutional networks being unable to capture temporal relationships and high training costs for 3D convolutional networks in traditional abnormal behavior detection. Firstly, a Time Shift Module (TSM) is added to the BoTNet backbone network to extract spatiotemporal feature information; Secondly, introduce the SimAM module to handle spatial and temporal dimensional features; Finally, softmax is used for classification. The experimental results showed that the accuracy in the HMDB51 dataset and UCF101 dataset reached 74.10% and 94.40%, respectively.
3. Using development platforms such as IntelliJ IDEA, this system adopts Java language to design and implement a human abnormal behavior detection system. Through testing and verification of actual data, this system has proven its effectiveness as an auxiliary monitoring tool in public place monitoring systems.