摘要
基于卷积神经网络(CNN)的目标检测方法已经在车辆检测中取得了巨大成功,然而自动场景下的车辆尺度变化较大,使得自动驾驶场景下的车辆检测依然是一个非常具有挑战性的问题,为此,针对车辆尺度变化这一问题提出了一种鲁棒的车辆检测方法。该算法在CNN不同尺度的特征图上提取目标候选框,进而将提取的这些候选框的特征输入另外一个小网络,对候选框做进一步的分类和回归。在通用的KITTI车辆检测数据集上的实验表明,方法可以大大提升基准算法检测的正确率,在1 280×384 pixels的自动驾驶场景的图像上,算法平均处理速度可以达到0.35 s/img。
Along with the success of convolutional neural network( CNN) in vehicle detection,the large variance in scales makes vehicle detection in automatic driving scenes a challenging task. In this paper,a robust vehicle detection method is proposed to handle the scale variance in automatic driving. This method utilizes feature maps of different levels in the CNN to generate object proposals of different scales,then these generated object proposals are fed into another tiny detection sub-network for further classification and regression. Experiments on the general KITTI vehicle object detection dataset show that the proposed method significantly outperforms the detection accuracy of the baseline,the processing speed can achieve 0. 35 s /img in the images of 1 280 × 384 pixels in the automatic driving scenes.
作者
范宜标
卢玮
傅智河
Fan Yibiao;Lu Wei;Fu Zhihe(School of Mechanical and Electrical Engineering,Longyan University,Longyan 364000,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2018年第12期60-65,共6页
Journal of Electronic Measurement and Instrumentation
基金
福建省自然科学项目(2017J01765)
龙岩学院青年攀登项目(LQ2017004)资助
关键词
自动驾驶
车辆检测
卷积神经网络
video surveillance
vehicle detection
convolutional neural network
