孙丽华,
徐志京
上海海事大学 上海 201306
基金项目:国家自然科学基金(61404083, 52071200)，中国博士后科学基金(2015M581527)，专用集成电路与系统国家重点实验室开放研究课题(2021KF010)
详细信息
作者简介:周薇娜：女，1982年生，副教授，研究方向为图像处理、电路和嵌入式系统、人工智能
孙丽华：女，1995年生，硕士生，研究方向为模式识别与图像处理
徐志京：男，1972年生，副教授，研究方向为海上智能交通系统、信息获取与智能处理
通讯作者:周薇娜 wnzhou@shmtu.edu.cn
中图分类号:TN911.73计量
文章访问数:337
HTML全文浏览量:136
PDF下载量:76
被引次数:0
出版历程
收稿日期:2020-06-01
修回日期:2020-12-01
网络出版日期:2021-03-31
刊出日期:2021-07-10
A Real-time Detection Method for Multi-scale Pedestrians in Complex Environment
Weina ZHOU,,Lihua SUN,
Zhijing XU
Shanghai Maritime University, Shanghai 201306, China
Funds:The National Natural Science Foundation of China (61404083, 52071200), China Postdoctoral Science Foundation (2015M581527), The State Key Laboratory of ASIC & System (2021KF010)
摘要
摘要:作为计算机视觉和图像处理研究领域中的经典课题，行人检测技术在智能驾驶、视频监控等领域中具有广泛的应用空间。然而，面对一些复杂的环境和情况，如阴雨、雾霾、被遮挡、照明度变化、目标尺度差异大等，常见的基于可见光或红外图像的行人检测方法的效果尚不尽如人意，无论是在检测准确率还是检测速度上。该文分析并抓住可见光和红外检测系统中行人特征差异较大，但在不同环境中又各有优势的特点，并结合多尺度特征提取方法，提出一种适用于多样复杂环境下多尺度行人实时检测的方法——融合行人检测网络(FPDNet)。该网络主要由特征提取骨干网络、多尺度检测和信息决策融合3个部分构成，可自适应提取可见光或红外背景下的多尺度行人。实验结果证明，该检测网络在多种复杂视觉环境下都具有较好的适应能力，在检测准确性和检测速度上均能满足实际应用的需求。
关键词:行人检测/
复杂环境/
自适应提取/
多尺度/
决策融合
Abstract:As a classic subject in computer vision and image processing, pedestrian detection has a wide range of applications to intelligence driving and video monitoring fields. However, most of pedestrian detection methods based on visible or infrared images have no satisfying result in some complex environments or situations, such as rain, smog, occlusion, variation of illuminance and target scales, no matter in terms of detection accuracy or speed. This paper analyzes and finds out that, pedestrians usually have quite different characteristics in visible and infrared image, and which have their own advantages in different environments. Therefore, combining fusion and multi-scale technology, a real-time multi-scale pedestrian detection algorithm suitable for complex environment named FPDNet (Fusion Pedestrian Detection Network) is proposed. The detection framework is consisted by three main modules: feature extraction backbone network, multi-scale detection network and decision-level fusion network. The proposed method is able to extract multi-scale pedestrian characteristics under visible or infrared background adaptively. Experimental results prove that the detection network has good adaptability in complex visual environments, and can meet the demands of practical applications to detection accuracy and speed.
Key words:Pedestrian detection/
Complex environment/
Adaptive extracting/
Multi-scale/
Decision-level fusion
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=31f5f021-5e1e-445b-83e0-626d277eb004
