自然场景下针对遮挡问题的目标检测算法研究  

Research on Object Detection Algorithm for Occlusion Problem in Natural Scenes

作　　者：王璐璐 朱荣 左云瑞 WANG Lulu;ZHU Rong;ZUO Yunrui(School of Mathematics and Computer Science,Hanjiang Normal University,Shiyan 442000,China;Yunnan Province Land Resources Planning and Design Institute,Kunming 650000,China)

机构地区：[1]汉江师范学院数学与计算机科学学院,湖北十堰442000 [2]云南省国土资源规划设计研究院,云南昆明650000

出　　处：《新乡学院学报》2025年第3期58-63,共6页Journal of Xinxiang University

基　　金：汉江师范学院科研基金资助课题(2024B06)。

摘　　要：针对自然场景下,背景复杂、小目标众多、排列密集导致的检测精度低、漏检和假阳性的问题,提出一种基于并行十字注意力机制的遮挡目标检测算法(PCA-OOD),该算法设计的水平最大池化和垂直最大池化能够获取显著的特征信息并减少计算量。随后构建出并行十字注意力机制,利用水平及垂直最大池化捕获各像素点水平及垂直方向的显著特征,以此来避免互相遮挡的同类别物体被误检,同时设计并行分支利用平均和水平池化捕获被主分支忽略的特征信息,降低漏检率。最后,在两大公开数据集PASCAL VOC和MS COCO中进行实验并验证检测器的有效性,获得了79.8%和43.6%的平均检测精确度。To solve the problems of low detection accuracy,frequent missed detections,and false positives resulting from complex backgrounds,numerous small targets,and dense arrangements in natural scenes,this paper proposes an occluded object detection algorithm based on a parallel cross attention mechanism(PCA-OOD).The horizontal and vertical maximum pooling designed in this algorithm can ob-tain significant feature information and reduce the computational burden.Subsequently,a parallel cross attention mechanism is constructed to capture the significant features of each pixel point in the horizontal and vertical directions using horizontal and vertical maximum pooling,thereby avoiding the misdetection of objects of the same category that are mutually occluded.Simultaneously,a paral-lel branch is designed to capture the feature information overlooked by the other branch using aver-age and horizontal pooling to reduce the missed detection rate.Finally,experiments are conducted in two major public datasets,PASCAL VOC and MS COCO,in this paper to verify the validity of the detector,and the average detection accuracies of 79.8%and 43.6%reobtained.

关 键 词：遮挡物体 目标检测 注意力机制 池化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]