改进Yolov5的无人机目标检测算法

UAV Target Detection Algorithm with Improved Yolov5

无人机场景下航拍图像存在密度高、目标小、覆盖范围广等特性,使得现有的目标检测器容易出现错检漏检的现象,为了提高识别的精度,提出了一种改进Yolov5的目标检测模型。通过采用梯度流丰富的C2F模块增加模型的特征提取能力。引入上采样算子CARAFE(content-aware reassembly of features)增加感受野进行数据特征融合,提升特征金字塔网络性能。通过采用全局性动态标签分配策略,提高模型识别准确率。通过VisDrone2019数据集验证表明,改进后的模型平均精度mAP值达到65.3%,较传统模型提升了24.7个百分点,可以更加准确地完成航拍过程中针对目标的检测任务。

The characteristics of high density,small target and wide coverage of aerial images in UAV scenes make the existing target detectors prone to error detection and omission,and in order to improve the recognition accuracy,a target detection model with improved Yolov5 is proposed.Firstly,the feature extraction capability of the model is increased by adopting the gradient flow-rich C2F module.Secondly,the up-sampling operator CARAFE(content-aware reassembly of features)is introduced to increase the perceptual field for data feature fusion and improve the feature pyramid network performance.Finally,the model recognition accuracy is improved by adopting a global dynamic label assignment strategy.The validation by VisDrone2019 dataset shows that the improved model achieves an average accuracy mAP value of 65.3%,which is 24.7 percentage points higher than the traditional model,and can more accurately complete the target-spe-cific detection task during aerial photography.