面向移动增强现实的实时深度学习目标检测方法综述

Review of real-time deep learning-based object detection for mobile augmented reality

移动增强现实(AR)借助智能移动终端将虚拟信息和真实世界进行实时融合,能否实时准确地对环境中需要增强的物体进行目标检测直接决定了系统的性能。随着深度学习的快速发展,近年来出现了大量的基于深度学习的目标检测方法。由于存在移动增强设备计算能力有限、能耗大、模型尺寸大以及卸载任务到边缘云端的网络延迟严重等问题,将深度学习方法应用于移动AR的目标检测是一项具有挑战性的问题。首先从Two stage和One stage的2方面对目前深度学习目标检测算法进行综述;然后对面向移动AR的目标检测系统架构进行归纳分类,分析了基于本地端、云端或边缘端和协作式的移动AR目标检测系统并总结了各自的优势和局限性;最后对移动AR中目标检测亟待解决的问题和未来发展方向进行了展望和预测。

Mobile augmented reality(AR)is a technology that integrates virtual information with the real world on the mobile intelligent terminal,therefore the ability to accurately detect the to-be-enhanced objects in the environment directly determines the performance of mobile AR systems.With the rapid advancement of deep learning,a large number of deep learning-based methods have been proposed for better detection.However,such problems as limited computing power,high energy consumption,large model size,and offloading latency make it difficult to combine deep learning-based object detection with mobile AR.This paper first summarized previous studies on deep learning-based object detection from both aspects of two stages and one stage,then categorized the object detection systems for mobile AR,and analyzed the approaches based on local,cloud,or edge ends,as well as collaboration.Finally,both the advantages and limitations of these methods were summarized,and predictions were made on the problems to be solved and the future development of object detection in mobile AR.