边缘环境下的视频行为识别

边缘智能是一种新兴的智能计算模式，其将人工智能技术和边缘嵌入式设备结合，被广泛应用于物联网系统。智能摄像机是典型的边缘设备之一，它能提供低延迟的视频处理能力，适用于智能家居、智能交通、智能监控等领域。然而，由于摄像机的计算资源有限，传统的行为识别模型难以在本地完成计算任务。为解决这一问题，文中提出了一种基于边缘计算的架构，利用深度学习目标检测算法YOLO v3对视频行为进行识别。在该架构中，智能移动终端负责数据采集和压缩，边缘服务器承担大部分目标检测任务，而检测困难的目标和模型训练则由云服务器负责。为更好地适应边缘设备，本文采用轻量化的神经网络MobileNet替换YOLO v3模型的特征提取模块。经过测试，该架构能有效提取和识别视频中的静态和动态行为，为实现边缘计算环境下低成本、大规模的行为识别提供了有益的参考。

Lightweight Video Behavior Recognition Method Based on Resnet

Edge intelligence is an emerging intelligent computing model, which combines artificial intelligence technology and edge embedded devices, and is widely used in Internet of Things systems. Smart cameras are one of the typical edge devices, which can provide low-latency video processing capabilities, which is suitable for smart homes, intelligent transportation, intelligent surveillance and other fields. However, due to the limited computing resources of cameras, traditional behavior recognition models are difficult to complete computing tasks locally. To solve this problem, this paper proposes an architecture based on edge computing, which uses deep learning object detection algorithm YOLO v3 to recognize video behaviors. In this architecture, intelligent mobile end points are responsible for data collection and compression, edge servers undertake most of the object detection tasks, while the detection of difficult targets and model training are handled by Cloud as a Service. In order to better adapt to edge devices, this paper uses lightweight neural networks MobileNet to replace the feature extraction module of YOLO v3 model. After testing, the architecture can effectively extract and identify static and dynamic behaviors in video, providing a useful reference for realizing low-cost and large-scale behavior recognition in edge computing environment.