基于形态检测与深度学习的高空视频车辆识别

为了准确提取广域场景道路交通信息，本文融合形态检测与深度卷积网络，提出了无人机视频车辆定位及车型识别方法. 首先，基于形态检测建立候选目标提取算法，并构建了含244 520 个无人机视频车辆样本的深度学习图像基准库；然后，通过增加卷积层、池化层及调整网络参数等方法对AlexNet 进行重构，提出了改进模型AlexNet*；最后，建立了基于候选目标提取算法与AlexNet*的车辆识别方法. 验证分析显示：AlexNet*的图像分类F1 均值达 85.51% ，优于AlexNet(82.54% )、LeNet(63.88% )、CaffeNet(46.64% )、VGG16(16.67% ) 及 GoogLeNet(14.38%)；本文车辆识别方法对小汽车及公交车的正检率、重检率和漏检率均值分别达94.63%、6.87%、4.40%，可有效识别无人机视频目标.

Morphous Detection and Deep Learning Based Approach of Vehicle Recognition in Aerial Videos

In order to extract traffic information accurately from a regional perspective, a vehicle locating and classifying method for UAV videos was proposed combining morphous detection and deep convolutional networks. Firstly, target candidate extraction algorithm was designed based on morphous detection, thus a benchmark image library for deep learning was constructed containing 244 520 vehicle samples from UAV videos. Then, AlexNet was rebuilt by adding convolutional and pooling layers and adjusting network parameters, etc., thus a modified model AlexNet* was put forward. At last, vehicle recognition method fusing candidate extraction and AlexNet* was established. Validation analysis indicate that AlexNet* achieves an average F1 score of 85.51% in terms of image classification, outperforming AlexNet (82.54%), LeNet (63.88%), CaffeNet (46.64%), VGG16 (16.67%) and GoogLeNet (14.38%), and the proposed vehicle recognition method averagely obtains accuracy of 94.63%, repeated detection rate of 6.87% and missing detection rate of 4.40% for car and bus detection, which can effectively identify UAV objects.