基本边缘计算的交通目标检测与识别

TRAFFIC TARGET DETECTION AND RECOGNITION BASED ON EDGE COMPUTING

詹鹏基

11849310

硕士

电子与通信工程领域工程

贡毅

2020-05-29

2020-07-20

哈尔滨工业大学

深圳

交通是社会生产和人类生活的重要组成部分，然而近年来随着机动车保有量的上升，交通引发的安全问题、环境破坏和经济损失日趋严重，人们开始使用智能交通系统（Intelligent Traffic System，ITS）解决问题。交通目标检测是智能交通必备的基本功能，一个实时性好、检测精度高的目标检测平台对于发展智能交通系统具有重要意义。传统目标检测方法由于检测精度太低早已被基于深度学习（Deep Learning，DL）的目标检测方法取代。然而，深度学习过高的计算复杂度极大地限制了检测速度。边缘计算（Edge Computing）致力于把计算任务和决策中心从云端转移到网络的边缘，以实现对时延敏感的应用场景的快速响应。本文提出了一种基于边缘计算的交通目标检测架构，结合深度学习检测算法Mask R-CNN，用于架构中的目标检测。文章首先对深度学习的理论知识和目标检测的工作原理进行介绍，然后对所提出架构中终端、边缘和云三个模块各自的任务进行设计：终端负责采集和压缩数据，边缘服务器负责大部分目标检测任务，云服务器负责一部分困难目标检测任务和所有目标检测模型的训练。同时对架构中云和边缘之间任务分配的置信度规则、边缘端目标检测模型的训练策略进行了超参数调优，达到效率与准确率之间的最佳平衡。对于嵌入边缘服务器的目标检测模型，本文使用轻量化神经网络MobileNet-V3作为模型的骨干网络以更好地适配边缘端的硬件环境，使用标准神经网络ResNet-50作为云服务器中模型的骨干网络以达到最优的检测性能。最后对所提出架构与基于云计算（Cloud Computing）的交通目标检测架构进行仿真比较。仿真结果表明，基于边缘计算的交通目标检测在仅牺牲少许检测精度（约5%mAP）的情况下，检测速度可以获得2倍以上的提升，对于网络带宽的占用只有后者的1/3。

Transportation is an important part of social production and human life. In recent years, with the increase in the number of motor vehicles, the safety problems, environmental damage and economic losses caused by transportation have become increasingly serious. People have begun to use Intelligent Traffic System (ITS) to solve problems. Traffic target detection is a basic function of intelligent transportation. A real-time object detection platform with high detection accuracy is of great significance for the development of intelligent transportation systems. Because of the low detection accuracy, the traditional object detection method has been replaced by the object detection method based on deep learning. However, the high computational complexity of deep learning greatly limits the detection speed. Edge Computing is aims to transfer computing tasks and decision centers from the cloud to the edge of the network to achieve rapid response to delay-sensitive application scenarios.In this paper, a traffic target detection architecture based on edge computing is proposed, which is combined with the deep learning detection algorithm Mask R-CNN for object detection. This paper first introduces the theoretical knowledge of deep learning and object detection, and then designs the respective tasks of the three modules of the terminal, edge and cloud in the proposed architecture. The terminal is responsible for collecting and compressing data, the edge server is responsible for most of the detection tasks, and the cloud server is responsible for a part of difficult object detection tasks and the training of all object detection models. The rules of the task allocation between the cloud and the edge and the training strategy of the object detection model of edge server are hyper-parametrically tuned to achieve the best balance between efficiency and accuracy. For the object detection model embedded in the edge server, this paper uses the lightweight neural network MobileNet-V3 as the backbone network of the model to better adapt to the hardware environment of the edge server, and uses the standard neural network ResNet-50 as the backbone network of the model in the cloud server to achieve the best detection performance. Finally, the proposed architecture is simulated and compared with the traffic target detection architecture based on cloud computing. The simulation results show that the detection speed of traffic target detection based on edge computing can be improved by more than 2 times at the expense of only a little detection accuracy (about 5%mAP), and the consumption of network bandwidth is only 1 / 3 of the latter.

边缘计算 深度学习 交通目标检测 Mask R-CNN

edge computing deep learning traffic target detection Mask R-CNN

中文

联合培养

南方科技大学

詹鹏基. 基本边缘计算的交通目标检测与识别[D]. 深圳. 哈尔滨工业大学,2020.