基于目标空间分布特征的无人机航拍车辆实时检测技术研究

针对无人机航拍视角下存在整体图像分辨率高但占比较高的小尺度车辆检测特征点稀少这一问题，从卷积网络检测器针对性优化与基于目标分布特征的航拍图像自适应切分2个角度综合考虑，提出一种基于目标空间分布特征的无人机航拍车辆检测网络DF-Net。以单阶段目标检测框架SSD为基础，引入深度可分离卷积和抗混叠低通滤波器对网络结构进行优化搭建E-SSD，为后续检测网络搭建提供高效检测器；接着基于条件生成对抗CGAN思想构建密度估计网络生成器，从而得到航拍图像中车辆的准确分布特征，生成高质量的车辆密度图；将E-SSD与车辆密度估计网络结合，对车辆密度图进行自适应切分，并将切分后的局部图像与全局图像一同输入E-SSD，最后在决策层融合检测结果，由此实现对航拍视角道路交通场景下车辆目标的精确高效检测。在试验中，一方面将设计的基于目标空间分布特征的无人机航拍车辆检测网络DF-Net与E-SSD进行对比分析，另一方面将DF-Net与航拍目标检测领域表现较为优秀的网络进行比较。研究结果表明：设计的方法对于2个试验在均值平均精度指标上均有提升，与E-SSD网络对比时提升了至少4.4%，与航拍目标检测领域优秀网络比较时也有一定提升，并保持了较好的实时性。

Real-time Vehicle Detection Technology for UAV Imagery Based on Target Spatial Distribution Features

To address the problem that the feature points of small-scale vehicles are sparse in high-resolution unmanned aerial vehicle (UAV) images, a UAV imagery vehicle detection network, DF-Net, is proposed from the perspectives of detector optimization and image-adaptive segmentation based on the target distribution characteristics. Our proposed detector is based on the single-phase target detection framework SSD, and the deeply separable convolution and anti-alias low-pass filter optimizes the network structure. This provides an efficient detector for subsequent detection network construction. Subsequently, a density estimation network generator is constructed based on the concept of the conditional generative adversarial nets (CGAN) to obtain the accurate distribution characteristics of vehicles in aerial images and generate a high-quality vehicle density map. The optimized efficient detector is combined with a vehicle density-estimation network. The vehicle density map is adaptively sliced, and the sliced local image is fed into the efficient detector together with the global image. Finally, the detection results are fused in the decision layer, thus achieving the accurate and efficient detection of vehicle targets in aerial photography views of road traffic scenes. In the experiments, the UAV imagery vehicle detection network DF-Net, which is based on the target distribution characteristics, was analyzed compared with the E-SSD, and it was also compared with networks that perform well in the field of aerial target detection. The experiment results show that the proposed method exhibits improved results for both experiments in terms of the mean average precision by at least 4.4% when compared with the E-SSD network. In addition, compared with networks in the field of aerial target detection, it maintains a better real-time performance.