多尺度特征增强融合的实时航拍车辆检测算法

Real-time vehicle detection in UAV photography based on multi-scale feature enhancement fusion

无人机航拍的车辆受成像条件及成像距离等限制,所成影像面临背景干扰强、尺度小的缺陷,传统的目标检测算法无法满足高精度和实时的车辆检测。本文提出多尺度特征增强融合的实时航拍车辆检测算法。首先,提出一种相邻特征连接模块,将浅层的细节特征和相邻深层的语义特征进行融合,形成多尺度特征金字塔模型。然后,设计一种通道特征增强模块,使得网络抑制冗余特征的学习,增强特征之间的判别性。实验证明,提出的航拍车辆检测算法能够达到91.3%的准确率和每秒58帧的实时性,可以较好解决处于遮挡、阴影干扰等复杂环境下的车辆弱小目标定位问题。

Due to the limitation of imaging conditions and distance,most vehicles by unmanned aerial vehicle(UAV)photography are of small size and accompanied by complex background.It is difficult for tra-ditional object detection methods to achieve high detection accuracy in real-time.This paper proposes a real-time vehicle detection method based on multi-scale feature enhancement and fusion.Firstly,we fuse shallow detailed features and adjacent deep semantic features using an adjacent feature connection module(AFCM)and form a multi-scale feature pyramid model.Secondly,we propose a channel feature enhancment module(CFEM)to suppress redundant features and improve the discriminability of target features.The experimental results show that the proposed UAV vehicle detection method can achieve 91.3%accuracy with a real-time speed of 58 frames per second.It also solves the problems of small vehicle detection in complex backgrounds such as occlusion and shadow interference.