基于改进YOLOv5的无人机航拍车辆检测算法

Drone Vehicle Detection Algorithm Based on Improved YOLOv5

摘要:

针对目前无人机航拍车辆检测中存在背景环境复杂、目标较小和计算复杂会造成漏检误检等问题,提出了一种改进YOLOv5的无人机车辆检测算法YOLOv5-R。采用轻量化模块GhostNetV2对主干网络进行更改,作为模型的主干特征提取网络,在缓解网络冗余的同时提高模型的检测速度；在主干网络中引入坐标注意力CA模块,增强模型对目标车辆检测的特征表达能力,从而提升模型在复杂背景下的检测精度；颈部引入加权双向特征金字塔网络BiFPN,增强模型的多尺度特征表达和融合能力,提升对小目标的检测精度；最后将原始的头部替换为动态检测头DyHead,通过大小、任务和空间感知的三者统一,进一步提高模型检测性能。实验结果表明,与原有的算法相比,改进YOLOv5算法的准确率和平均精确度分别提高了6.5%和5.1%,且算法检测速度达到99.7 FPS,满足检测实时性的要求,与其他主流模型相比,该模型在公开数据集上有更好的检测效果,验证了其可行性和有效性。

Abstract:

A novel unmanned aerial vehicle detection algorithm YOLOv5-R is proposed to address the issues of complex background environments, small targets, and complex calculations that can lead to missed detections and false detections in current drone aerial vehicle detection. Using the lightweight module GhostNetV2 to modify the backbone network as the backbone feature extraction network of the model, while alleviating network redundancy and improving the detection speed of the model; Introducing a coordinate attention CA module into the backbone network enhances the model"s feature representation ability for target vehicle detection, thereby improving the detection accuracy of the model in complex backgrounds; Introducing a weighted bidirectional feature pyramid network BiFPN into the neck enhances the model"s multi-scale feature representation and fusion capabilities, and improves the detection accuracy of small targets; Finally, the original head is replaced with the dynamic detection head DyHead, which further improves the detection performance of the model by unifying size, task, and spatial perception. The experimental results show that compared with the original algorithm, the improved YOLOv5 algorithm improves accuracy and average accuracy by 6.5% and 5.1%, respectively, and the detection speed of the algorithm reaches 99.7 FPS, meeting the requirements of real-time detection. Compared with other mainstream models, the model has better detection performance on public datasets, verifying its feasibility and effectiveness.