摘要
针对大范围快速的车辆检测与计数,利用高分辨率卫星影像数据,提出了一种基于区域卷积神经网络的车辆检测算法。区域卷积神经网络是深度卷积神经网络和区域建议网络二者的结合。首先利用深度卷积神经网络自动提取各个层的特征,为了减少检测窗口的数量,提出区域建议网络,对下采样后的每个位置考虑3种窗口和对应的3种比例,这样大大减少了检测窗口的数量。再根据分类器对目标进行分类。这样将特征、检测窗口和分类器有效地结合在一起。在对遥感影像车辆检测试验中,通过对手工标注的车辆样本数据多次迭代来训练卷积神经网络和区域建议网络获取车辆检测的先验模型,再由先验模型检测出测试影像中车辆目标。与传统的基于梯度方向直方图(HOG)特征和支持向量机(SVM)车辆检测算法进行比较,在检测率方面,区域卷积神经网络算法明显高于HOG+SVM算法;在误检率方面,区域卷积神经网络检测明显小于HOG+SVM算法;在检测时间方面,同样的一张图像,区域卷积神经网络检测速度比HOG+SVM算法提升近800倍。试验结果表明:利用区域卷积神经网络方法进行大范围车辆检测,在精度和速度方面都有显著提升。
For a wide range and fast vehicle detection and counting, a vehicle detection algorithm based on regional convolution neural network is proposed based on high resolution satellite image data. Regional convolution neural network is the integration of deep convolution neural network and region proposal network. First, deep convolution neural network is used to automatically extract the features of each layer. In order to reduce the number of detection windows, a regional suggested network is proposed. Three windows and 3 corresponding proportions are considered for each location after down-sampling, which greatly reduced the number of detection windows. Then, the targets are classified according to the classifier. It effectively combines features, detection windows and classifiers. In the experiments of vehicle detection with remote sensing images, a prior model of vehicle detection is obtained by using the multi-iteration of manual labeled vehicle sample data to train the convolution neural networks and the region proposal networks. Then, vehicles are detected in tested remote sensing images by this model. Compared with the traditional algorithm based on gradient direction histogram （HOG） and support vector machine （SVM）, the result of the regional convolution neural network algorithm is obviously higher than that of the HOG ＋ SVM algorithm in the detection rate, the resuh of the regional convolution neural network is obviously smaller than that of the HOG ＋ SVM algorithm in the error detection rate, the detection speed of the region eonvolutinn neural network is nearly 800 times faster Ihan that of the HOG ＋ SVM algorithm for the same image in detection time. The experimental resuh shows that the method of the region convolution neural network has a significant improvement in the accuracy and speed for a wide range of vehicle detection.
出处
《公路交通科技》
CAS
CSCD
北大核心
2018年第3期103-108,共6页
Journal of Highway and Transportation Research and Development
基金
国家自然科学基金项目(41372330
41601345)
关键词
交通工程
车辆检测
卷积神经网络
区域建议网络
遥感影像
traffic engineering
vehicle detection
conw）lutional neural network
region proposal network
remote sensing image
