摘要： 边缘环境下的目标检测应用大多依赖于边缘设备。目前，MobileNet等轻量级检测模型在满足实时性要求的前提下难以达到准确度要求，Faster RCNN等重量级模型的检测准确度较高，但数据传输耗时大，难以保证检测实时性。提出一种边缘辅助的目标检测框架及自适应帧卸载（OEOD）算法，在线将视频帧卸载到边缘服务器，以执行高准确度、高时延的目标检测，或在本地执行低准确度、低时延的目标检测，从而在满足给定检测准确度要求的前提下缩短视频帧的平均处理时延。为实现OEOD算法，提出一种基于内容相似性的特征向量度量方式，以预测当前帧的检测时延和准确度，并以贪心的方式决定当前帧的检测策略。实验结果表明，与MobileNet算法相比，OEOD算法在保证准确度的前提下将检测时延降低了29%，且在不同的数据集上均表现良好。

关键词: 边缘计算, 目标检测, 计算卸载, 在线调度, 时延优化

Abstract: The application of object detection in edge environment mostly depends on the edge device. At present，lightweight detection models，such as MobileNet，have difficulty simultaneously meeting real-time and accuracy requirements. Heavyweight models，such as Faster RCNN，have high detection accuracy but data transmission is time-consuming，making it difficult to ensure real-time detection. An edge assisted object detection framework and adaptive frame offloading algorithm called OEOD is proposed，which，given an accuracy requirement，adaptively chooses to offload a video frame to an edge server for high-accuracy high-delay object detection or locally run low-accuracy low-delay object detection，which optimizes the average per-frame detection delay. To implement OEOD algorithm，we also propose a feature vector measurement method based on content similarity to predict the detection delay and accuracy of the current frame and determine the detection strategy of the current frame in a greedy manner. Experimental results show that OEOD algorithm can reduce the per-frame delay by 29% compared with MobileNet given a fixed accuracy requirement，and performs well on different datasets.

Key words: edge computing, object detection, computation offloading, online scheduling, delay optimization

中图分类号: 

• TP391