摘要
交通场景下的目标因为密集或遮挡等情况会导致检测困难,针对该问题,提出了一种交通场景目标检测算法FFE-YOLOv8。设计跨层级特征融合模块(cross-level feature fusion module,CFFM),结合浅层细节和深层整体特征,得到更丰富的特征信息指导遮挡目标的检测;设计了双路特征增强模块(dual-channel feature enhancement,DCFE),经两通道同时提取浅层特征的中心区域关键信息以及上下文信息,补充特征的同时避免浅层特征层感受野较小的问题;经自适应特征融合模块(adaptive feature fusion module,AFFM),计算经双路特征补充模块处理后的特征和浅层检测通道特征的相似度从而进行有针对性的增强。在KITTI,CBCL和Udacity数据集上进行实验,结果显示相比于基准网络改进后的算法在精度上分别提升了1.52、1.61、2.47个百分点,改善了遮挡目标的检测效果,证明了该方法的有效性。
Objects in traffic scenes are difficult to detect because of density or occlusion.To solve this problem,a traffic scenes object detection algorithm FFE-YOLOv8 is proposed.Firstly,a cross-level feature fusion module(CFFM)is designed.It combines shallow details and deep features to obtain richer feature information for guiding occlusion detection.Secondly,a dual-channel feature enhancement module(DCFE)is designed.It simultaneously extracts the key information of the central region and the context information of the shallow feature through two channels.This avoids the problem of the small receptive field of the shallow feature layer while supplementing the feature.Finally,an adaptive feature fusion module(AFFM)is used to calculate the similarity between the features processed by the dual-channel feature enhancement mod-ule and those of the shallow detection channel.Experiments on the KITTI,CBCL and Udacity datasets show that com-pared with the benchmark network,the improved algorithm accuracy is increased by 1.52,1.61 and 2.47 percentage points respectively.The occlusions detection effect is improved,which proves the effectiveness of this method.
作者
王坤
冯康威
WANG Kun;FENG Kangwei(College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)
出处
《计算机工程与应用》
北大核心
2025年第15期156-166,共11页
Computer Engineering and Applications
基金
国家自然科学基金(62173331)。
