基于改进YOLOX的多光谱行人检测算法  

Multispectral Pedestrian Detection Based on Improved YOLOX

作　　者：方康 黄琴 王克琪 靳帅 刘畅 钱宇华 陈路[1,3,4] FANG Kang;HUANG Qin;WANG Keqi;JIN Shuai;LIU Chang;QIAN Yuhua;CHEN Lu(Institute of Big Data Science and Industry,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing Ministry of Education,Shanxi University,Taiyuan 030006,China;School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Taiyuan Satellite Launch Center,Taiyuan 030012,China)

机构地区：[1]山西大学大数据科学与产业研究院,太原030006 [2]山西大学计算智能与中文信息处理教育部重点实验室,太原030006 [3]山西大学计算机与信息技术学院,太原030006 [4]太原卫星发射中心,太原030012

出　　处：《小型微型计算机系统》2024年第1期185-191,共7页Journal of Chinese Computer Systems

基　　金：国家重点研发计划项目(2021ZD0112400)资助;国家自然科学基金项目(62136005,62106132,62003200)资助.

摘　　要：现有的多光谱行人检测算法大多是基于Faster R-CNN的两阶段检测或设置了锚框机制的一阶段检测,此类模型存在推理速度慢,检测准确率低等不足.为此,本文设计出一种基于一阶段无锚框检查算法YOLOX的多光谱行人检测算法.该算法将多模态特征提取解耦为特性特征提取和共性特征提取两阶段.针对基准特性特征提取网络学习能力不足、提取的语义信息和纹理细节信息不够丰富的问题,本文设计出一种多尺度特征增强(Multi-scale Feature Enhancement,MFE)模块,该模块可提取出特性特征图中丰富的语义和纹理细节信息.此外,本文使用基于差异性的特征融合方法来充分捕获两种模态的差异性特征信息.为证实本文方法的有效性和可行性,本文在KAIST数据集和FLIR数据集上进行了实验验证,实验结果表明本文所提方法可显著提高多光谱行人检测的性能.Most of existing multispectral pedestrian detection algorithms are based on two-stage Faster-RCNN or one-stage detection algorithms with anchor boxes,which suffer from low inference speed and low detection accuracy.To handle this,we design an one-stage anchor-free method which uses YOLOX as the detector with two feature extraction modules:modality-specific feature extraction and modality-common feature extraction.Traditional modality-specific feature extraction modules have defects in learning capacity,mining detailed semantic and texture information.In this paper,we tackle the above limitation by proposing a novel Multi-scale Feature Enhancement module(MFE),it can extract abundant semantic and textureinformation.Furthermore,for capturing the different characteristic of two modalities,we use the specific fusion strategy which is based on differences.To verify the effectiveness and feasibility of our method,we conducted comprehensive experiments on KAIST and FLIR datasets to demonstrate superiority of the overall algorithm and the effectiveness of each component.The proposed algorithm significantly improves the detection accuracy.

关 键 词：多光谱 行人检测 多尺度 YOLOX 一阶段 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]