空对地场景信息辅助的目标检测算法

Target Detection Algorithm Using Air-to-Ground Scene Information

在空对地视角下图像的场景信息往往更加丰富,并且对目标的定位和分类有很强的辅助作用.传统的单发多框检测(SSD)网络在6个不同深度的特征图上对目标边框和类别独立地进行预测,忽略了深层次特征的场景信息对浅层细节信息的辅助作用.为有效地利用场景信息,首先在SSD网络的基础上分析不同尺度的特征图对目标检测的影响;然后结合特征金字塔和长短期记忆网络针对不同特征图建立场景辅助结构,增强特征图的表征能力.在自制的空对地数据集上进行实验,并与检测领域几种经典的网络进行对比,结果表明,文中算法能够在保证速度的前提下高效地对目标进行检测,比其他经典的网络有更高的检测精度.

The scene information of image is more abundant in the air-to-ground perspective, which is helpful for the location and classification of target. The traditional single shot multibox detector(SSD) net predicts bounding boxes and classifications of the target from six feature maps, ignoring the auxiliary effect of high-level semantic feature maps on the ?ne details of shallow layers. In order to combine scene information, firstly, the effects of different scale feature maps on target detection are analyzed on the basis of SSD net. And then a new scene auxiliary structure is established by combining the ideas of feature pyramid networks and long short term memory net to enhance representation of feature maps. At last, the method is validated on a self-made dataset of air-to-ground and compared with several classic networks in the detection domain. The results show that the proposed method has higher detection accuracy than others, which can detect targets efficiently under the prerequisite of guaranteeing speed.