摘要
针对行人翻越护栏检测在遮挡、多目标密集以及多人翻越的复杂场景下出现的漏检、误检和检测精度低的问题,提出了一种多尺度特征融合增强的行人翻越护栏检测算法。首先,设计了一种基于Dual Vision Transformer和SCConv构建的SCDVT模块,应用于主干网络,增强了对全局上下文信息和更细粒度信息的捕获,提升了网络的局部精细特征提取和特征融合能力;其次,提出多尺度特征融合增强模块AM-SPPFCSPC,弥补了最大池化带来的特征损失,提高了特征图的丰富性和完整性,增强了多尺度特征提取和特征融合能力;另外,对特征融合层再进行细化,使用GSConv替换普通卷积,并基于GSConv和SCConv设计了VOV-GSCCSP模块,有效的降低了计算成本和模型的复杂度,同时又保持较高的精度;最后在主干引入高效多尺度注意力EMA,减少了复杂背景下无关目标的干扰,融合了多尺度信息,实现了更丰富的特征聚合。在自制行人翻越护栏数据集上的实验结果表明,本文所提算法在增加较少参数量的情况下,其mAP达到了93.6%,较原模型提高了4.5%,并且检测速度为108.5 FPS,改善了漏检、误检和检测精度低的问题,同时仍具有较高的实时性,更适用于行人翻越护栏的实时性检测。
Aiming at the problems of omission,misdetection and low detection accuracy of pedestrian crossing guardrail detection in the complex scenarios of occlusion,dense multi-target situations as well as multiple people overtopping,a multi-scale feature fusion and enhancement algorithm for pedestrian overtopping guardrail detection is proposed.Firstly,an algorithm based on Dual Vision Transformer and SCConv,which is applied to the backbone network,enhances the capture of global context information and finer-grained information,and improves the local fine feature extraction and feature fusion capability of the network;second,a multi-scale feature fusion enhancement module AM-SPPFCSPC is proposed,which compensates for the feature loss caused by maximal pooling,improves the feature map.The richness and completeness of the feature map is improved,and the multi-scale feature extraction and feature fusion capability is enhanced;in addition,the feature fusion layer is further refined by replacing the ordinary convolution with GSConv and designing the VOV-GSCCSP module based on GSConv and SCConv,which effectively reduces the computational cost and the complexity of the model,while maintaining a higher degree of accuracy;finally,a highly efficient multi-scale feature fusion module,AM-SPPFCSPC,is introduced in the trunk to reduce the complex background and the complex background and the complex background and the complex background.Attention EMA,which reduces the interference of irrelevant targets in the complex background and fuses the multi-scale information to achieve richer feature aggregation.The experimental results on the homemade pedestrian over guardrail dataset show that the proposed algorithm in this paper achieves 93.6%mAP with the addition of fewer parameters,which is 4.5%higher than that of the original model,and has a detection speed of 108.5 FPS,which improves the problems of leakage,false detection and low detection accuracy,while still having a high real-time performance,and is more suitable for real-time detection of pedestrians crossing the guardrail.
作者
刘罡
侯恩翔
黄孙港
闫曙光
黄应征
Liu Gang;Hou Enxiang;Huang Sungang;Yan Shuguang;Huang Yingzheng(Jiangsu Province Engineering Research Center of Integrated Circuit ReliabilityTechnology and Testing System,Wuxi University,Wuxi 214105,China;School of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China)
出处
《电子测量技术》
北大核心
2024年第14期127-138,共12页
Electronic Measurement Technology
基金
国家自然科学基金(62204172)项目资助。
