摘要
在真实的交通场景中,车辆和行人检测器容易受到复杂背景的干扰,导致严重的误检和漏检。此外,由于交通场景图像中存在各种不同尺度的目标,大多数检测方法对小物体的检测性能较差,且难以适应交通物体的多样性。为了解决这些问题,提出了一种基于注意力和前景注意模块的无锚方法VHA-CenterNet。在主干网络中加入一个卷积块注意模块(CBAM),以提高对小目标的关注能力。在前景注意模块(FAM)中引入前景信息,以减少复杂背景的干扰。结果表明:在中等难度下,所提出的VHA⁃CenterNet方法在KITTI数据集上的mAP达71.92%,在RTX 2080 Ti上的推理速度为10.68 FPS,可以显著提高人车识别的准确率和速度。在所有情况下,交通场景的人车检测准确率都高于传统模型。
In real traffic scenarios,vehicle and pedestrian detectors are susceptible to interference from com⁃plex backgrounds,leading to serious false and missed detections.Additionally,most detection methods have poor detection performance for small objects and are difficult to adapt to the diversity of traffic objects due to the presence of various targets of different scales in the traffic scene images.To address these problems,an ef⁃ficient anchor-free method called VHA-CenterNet is proposed based on attention and foreground attention modules.Firstly,a convolutional block attention module(CBAM)is added to the backbone network to im⁃prove the ability to focus on small targets.Secondly,the foreground attention module(FAM)introduces fore⁃ground information and reduces the interference of complex backgrounds.The results show that at moderate dif⁃ficulty,the VHA-CenterNet method proposed achieves a mAP of 71.92%on the KITTI dataset and an infer⁃ence speed of 10.68 FPS on RTX 2080 Ti,which can significantly improve the accuracy and speed of the hu⁃man-vehicle recognition.The accuracy of the human-vehicle detection for traffic scenes is higher than that of the traditional model in all cases.
作者
陈鑫影
吕娜
吕硕
CHEN Xinying;LYU Na;LYU Shuo(School of Computer and Communication Engineering,Dalian Jiaotong University,Dalian 116028,China)
出处
《大连交通大学学报》
CAS
2024年第4期101-107,共7页
Journal of Dalian Jiaotong University
基金
辽宁省应用基础研究计划项目(1655706734383)。
关键词
图像处理
人车检测
深层聚合
深度学习
注意力
image processing
human-vehicle detection
deep layer aggregation
deep learning
attention
