基于渐进式多光谱图像的行人检测算法  

Pedestrian detection algorithm based on progressive multispectral images

作　　者：张波 赵云鹏[1,2] ZHANG Bo;ZHAO Yun-peng(Department of Computer Science and Technology,Shenyang University of Chemical Technology,Shenyang 110142,China;Liaoning Key Laboratory of Industrial Intelligence Technology on Chemical Process,Shenyang University of Chemical Technology,Shenyang 110142,China)

机构地区：[1]沈阳化工大学计算机科学与技术学院,辽宁沈阳110142 [2]沈阳化工大学、辽宁省化工过程工业智能化技术重点实验室,辽宁沈阳110142

出　　处：《计算机工程与设计》2024年第10期3042-3050,共9页Computer Engineering and Design

基　　金：辽宁省博士科研启动基金项目(2019-BS-191);辽宁省教育厅科学研究基金项目(LJ2020023)。

摘　　要：针对现有单光谱图像的行人检测算法在全天候环境下工作效率差的问题,提出一种基于渐进式多光谱图像的行人检测算法。设计一种差分位置注意增强模块,嵌入到不同层次的双流CSP-Darknet53特征主干提取网络中,渐进式增强不同光谱之间的全局互补信息;采用跨模态互补信息融合策略对现有融合方法进行改进,利用全局特征引导多尺度特征融合,进行更鲁棒性的行人检测。实验结果表明,该算法在LLVIP数据集上mAP50值为97.2%,在FLIR数据集上获得84.6%的mAP50值,具有良好的检测性能。In view of the poor efficiency of existing pedestrian detection algorithms based on single-spectral images in all-weather environments,a pedestrian detection algorithm based on progressive multispectral images was proposed.A differential position attention enhancement module was designed,which was embedded in the dual-stream CSP-Darknet53 feature backbone extraction network at different levels,to incrementally enhance global complementary information between different spectra.The cross-modal complementary information fusion strategies were used to improve existing fusion methods,global features were used to guide multi-scale feature fusion for more robust pedestrian detection.Experimental results show that the mAP50 value on the LLVIP dataset is 97.2%,and the mAP50 value on the FLIR dataset is 84.6%,indicating good detection performance.

关 键 词：行人检测 多光谱图像 跨模态融合 注意力机制 融合策略 多尺度特征 全天候场景 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]