Improved YOLOV3 Surveillance Device Object Detection Method Based on Federated Learning

Abstract

In the process of smart city construction, a large number of video monitoring devices are used for urban security. However, these devices produce a large amount of monitoring data, which will make the task of retrieving specific goals difficult to achieve. At the same time, as people attach importance to privacy protection, it is difficult for data interaction between different enterprise devices. In this paper, in order to speed up the operation efficiency and break the data island of the IoT devices, we make the edge device of video surveillance for object detection based on federated learning, and improve YOLOv3. In order to simplify the model, Resnet50 is used to replace Darkent-53. We propose Multi-Dimensional and Multi-Scale (MDMS) feature pyramid to better use features to detect objects at different scales. We design the Improved OutLook Attention (IOA) module to make the network perform better in fine-grained feature. The final experimental results show that our network has greatly improved video surveillance object detection. Our method improves the mapping of the model and reduces the number of model parameters. The model size is compressed from 469.64 MB to 180.39 MB. However, the map is increased from 83.5% to 85.8%, which makes the video monitoring equipment more intelligent and enhances the ability of privacy protection.