基于改进YOLOv5s的轻量级无人机遥感影像车辆检测方法

Lightweight UAV Remote Sensing Image Vehicle Detection Method Based on Improved YOLOv5s

针对无人机机载硬件条件下的地面车辆目标实时检测问题,提出一种基于YOLOv5s的轻量级检测模型。模型在骨干网络中引入了分组卷积核进行特征提取,来降低骨干网络的运算参数量,同时在每个特征提取层末端设置了高效注意力机制来提高对正样本特征的赋权筛选,在特征增强端前方设置了多尺度特征融合层,进一步提高输出特征图内的信息丰富程度。实验结果表明,提出的改进模型在检测精度、速度以及模型体量方面均优于原始模型,在夜景、高速公路等多种复杂场景下表现出了良好的泛化能力,能够部署于无人机机载硬件中开展车辆目标的实时检测。

This paper proposes a lightweight detection model based on YOLOv5s for the real-time detection of ground vehicle targets under the condition of UAV airborne hardware.The model introduces a grouped convolution kernels into the backbone network for feature extraction to reduce the number of operational parameters of the backbone network,and at the same time,an efficient attention mechanism is set at the end of each feature extraction layer to improve the weighting and screening of positive sample features.A multi-scale feature fusion layer is set in front of the feature enhancement end to further improve the information richness in the output feature map.The experimental results show that the proposed improved model is superior to the original model in terms of detection accuracy,speed and model volume.It can show good generalization ability in various complex scenes such as night scenes and highways and can be deployed in UAV airborne hardware to carry out real-time detection of vehicle objects.