一种基于帧间运动目标框链的视频目标检测算法

李敏; 李凌涵; 白入文; 姜淼; 任俊星; 孟博; 杨阳; 黄子豪; 黄伟庆

引用本文：

李敏,李凌涵,白入文,姜淼,任俊星,孟博,杨阳,黄子豪,黄伟庆.一种基于帧间运动目标框链的视频目标检测算法[J].信息安全学报,2026,11(3):238-248 [点击复制]
LI Min,LI Linghan,BAI Ruwen,JIANG Miao,REN Junxing,MENG Bo,YANG Yang,HUANG Zihao,HUANG Weiqing.A Motion-based Seq-bbox Matching Method for Video Object Detection[J].Journal of Cyber Security,2026,11(3):238-248 [点击复制]

本文已被：浏览 139次下载 167次	码上扫一扫！
一种基于帧间运动目标框链的视频目标检测算法
李敏^1,2, 李凌涵^1,2, 白入文^1,2, 姜淼^1,2, 任俊星^1,2, 孟博³, 杨阳^1,2, 黄子豪⁴, 黄伟庆^1,2
字体:加大+\|默认\|缩小-
(1.中国科学院信息工程研究所北京中国 100093;2.中国科学院大学网络空间安全学院北京中国 100049;3.北京理工大学北京中国 100081;4.华北电力大学北京中国 102206)

摘要:

目标检测是计算机视觉领域的基础任务,在视频监控、自动驾驶等领域有着广泛的应用。近年来,随着深度卷积神经网络的发展,YOLO、Faster R-CNN等基于单帧图像的目标检测算法已经取得显著的效果,但在实际使用中,由于视频中目标物体模糊、遮挡、变形等问题,单帧图像目标检测算法无法直接应用在视频目标检测任务中。目前主流的视频目标检测算法,如FGFA、Seq-NMS等,无法同时兼顾速度与准确性,不能满足实际应用的需求。如何在保障检测实时性的前提下,提升检测的准确性,是当前视频目标检测算法面临的挑战。本文提出一种基于帧间运动目标框链的后处理方法,以单帧检测算法为基础,从后处理部分入手,引入帧间运动信息对检测结果进行增强。我们利用基于距离的交并比(Distance Intersection over Union,DIoU)表示帧间运动信息,并提出相邻帧之间的同一个物体应该具有相似的运动信息的观点,完成帧间运动目标框链的构建。最后,引入动态置信度平均方法,来对预测结果进行增强。实验结果表明,以YOLOv5为基础,本算法达到了73.4%的平均精度均值(mean average precision,mAP),获得了6.2%(67.2%~73.4%)的mAP提升,同时检测速度达到了41每秒处理帧数(frames per second,fps)。本算法很好地兼顾速度和准确度,并为如何进行更加快速准确的视频目标检测算法提供思路。

关键词: 视频目标检测检测框运动信息基于距离的交并比

DOI：10.19363/J.cnki.cn10-1380/tn.2026.05.16

投稿时间：2021-03-09修订日期：2021-04-26

基金项目:本课题得到国家重点研发计划(No.2019YFB1005205)资助。

A Motion-based Seq-bbox Matching Method for Video Object Detection

LI Min^1,2, LI Linghan^1,2, BAI Ruwen^1,2, JIANG Miao^1,2, REN Junxing^1,2, MENG Bo³, YANG Yang^1,2, HUANG Zihao⁴, HUANG Weiqing^1,2

(1.Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100093, China;2.School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China;3.Beijing Institute of Technology, Beijing 100081, China;4.North China Electric Power University, Beijing 102206, China)

Abstract:

Object detection is a fundamental task in the field of computer vision and is widely applied in video surveillance, autonomous driving, robotics and other fields. In recent years, with the development of deep convolutional neural networks, single-frame image-based object detection algorithms, such as YOLO and Faster R-CNN, have achieved remarkable results. However, in practice, single-frame image-based object detection algorithms cannot be directly applied to video object detection tasks due to the problems of blurring, occlusion, and deformation of the target objects in the video. The current mainstream video object detection algorithms, such as FGFA, Seq-NMS, etc., are unable to simultaneously take into account the detection speed and accuracy, thus cannot meet the needs of practical applications. How to improve the accuracy of detection under the premise of guaranteeing the real-time detection is the challenge faced by the current video object detection algorithms. In this paper, we propose a post-processing method based on inter-frame motion object frame chain. Based on the single-frame object detection algorithm, the video object detection results are enhanced by introducing inter-frame motion information from the post-processing part. We utilize Distance Intersection over Union (DIoU) to represent the inter-frame motion information, and propose the idea that the same object between adjacent frames should have similar motion information. We then combine the dynamic confidence averaging method to jointly complete the enhancement of the prediction results. Finally, the dynamic confidence averaging method is introduced to enhance the prediction results. The experimental results show that, based on YOLOv5, the proposed algorithm achieves a mean average precision (mAP) of 73.4% and obtains a mAP enhancement of 6.2% (67.2%-73.4%). At the same time, the detection speed reaches 41 frames processed per second (fps). The proposed video object detection algorithm achieves excellent results in terms of balancing speed and accuracy and provides ideas on developing faster and more accurate video object detection algorithms.

Key words: video object detection bounding box motion information Distance Intersection-over-Union