改进YOLOv5的无人机影像道路目标检测算法

Road target detection algorithm based on improved YOLOv5 in UAV images

针对无人机影像中道路小目标漏检和目标之间遮挡导致的目标检测精度低、鲁棒性差等问题,提出一种多尺度融合卷积注意力模块(Convolutional block attention module,CBAM)的YOLOv5道路目标检测算法,即YOLOv5s-FCC。首先,引入小目标感知层对模型进行多尺度改进,增加一个针对小目标的YOLO检测头以提高网络对道路中小目标的特征提取能力。其次,利用CBAM融合空间和通道信息以增强网络中的重要信息,通过将CBAM引入Backbone主干网络不同位置,以获得CBAM最佳融合位置。最后,采用CIo U作为损失函数,以提高边界框预测所需的计算速度和精度。在自建的无人机道路目标数据集上进行训练,结果表明,相较YOLOv5算法,YOLOv5-FCC算法可将mAP50和mAP50-95分别提高2.0%和4.2%。在VisDrone数据集上也验证了YOLOv5-FCC算法的有效性,并建立了基于无人机的道路目标检测系统,实现道路目标的自动检测。

Aiming at the problems such as low accuracy and poor robustness of target detection caused by missed detection of small road targets and occlusion between targets in UAV images,an improved road target detection algorithm based on YOLOv5 combining convolutional block attention module(CBAM),called YOLOv5s-FCC,was proposed.Firstly,a small target sensing layer was introduced to improve the multi-scale model,and a small target YOLO detection head was added to improve the feature extraction ability of the network for small road targets.Secondly,the CBAM fused space and channel information to enhance important information in the network after it was introduced into different locations of the Backbone network to obtain the best fusion location of CBAM.Finally,CIoU loss function was used to improve the speed and accuracy of the calculation required for predicting the bounding box of image.The experimental results showed that compared with YOLOv5 algorithm,YOLOV5-FCC algorithm can improve mAP50 and mAP50-95 by 2.0% and 4.2%,respectively.The effectiveness of YOLOv5-FCC algorithm was also verified on VisDrone dataset,and the results showed that the established system can realize automatic detection of road targets.