摘要
针对驾驶场景下的行人检测面临的环境复杂、行人密集和尺度跨越大等问题,提出一种智能驾驶场景下的端到端行人检测方法。为减少特征金字塔直接对特征相加造成的信息损失,引入双向特征增强模块(Bidirectional Feature Enhancement Module,BFEM),在双向通道上使用级联融合增强各层特征包含的信息。针对检测器在行人遮挡场景下感知力不足的问题,提出一种注意力部位感知模块(Embedding-based Attention Part-aware Module,EAPM),模块使用任务感知注意力增强特征前景特性,同时为人体部位添加了可见性损失,以此来增强模型对人体结构的感知经验。此外,改进任务感知注意力结合空间分组思想,增强子特征信息,减少噪声干扰,以此增强检测器的分类能力。在CrowdHuman和Citypersons数据集上对模型进行评估,实验证明了方法的有效性,在CrowdHuman中与基线相比提升了2.39%的AP值、2.21%的Recall和3.08%的R_(M)^(-2)值,取得了91.55%AP,89.88%Recall和43.90%R_(M)^(-2)的结果,在Citypersons中取得了44.4R_(M)^(-2)的结果。
Pedestrian detection in driving scenarios faces challenges such as complex environments,dense pedestrian population and large scale span.This paper proposes an end-to-end pedestrian detection method for intelligent driving scenarios.To address the scale issue and reduce information loss caused by directly adding features from the feature pyramid,a Bidirectional Feature Enhancement Module(BFEM)is introduced,which enhances the information contained in each layer of features through concatenated fusion and bi-directional channels.To address the problem of insufficient perception of the detector in pedestrian occlusion scenes,this paper adds the Embedding-based Attention Part-aware Module(EAPM)to the detector,which uses task-aware attention-enhancing feature foreground features,while adding visibility loss for human parts as a way to enhance the model's perceptual experience of the human structure.In addition,this paper improved the idea of task perceptual attention combined with spatial grouping,enhanced sub-feature information,reduced noise interference,and thus enhanced the classification ability of the detector.The proposed method is evaluated on the CrowdHuman and Citypersons dataset,and the experimental results demonstrate its effectiveness.Compared with the baseline,it achieves significant improvements of 2.39%in AP,2.21%in Recall,and 3.08%in R_(M)^(-2),achieving 91.55%AP,89.88%Recall,and 43.90%R_(M)^(-2).On Citypersons dataset,it achieves a result of 44.4R_(M)^(-2).
作者
詹智祺
程艳云
ZHAN Zhiqi;CHENG Yanyun(College of Automation&College of Artificial Intelligence,Nanjing University of Posts and Telecommunications,Nanjing 210046,China)
出处
《微电子学与计算机》
2024年第8期31-39,共9页
Microelectronics & Computer
基金
国家自然科学基金青年科学基金(62001247)。
关键词
行人检测
特征融合
端到端目标检测
部位感知
任务感知注意力
pedestrian detection
feature fusion
end-to-end object detection
part-aware
task-aware attention
