基于多尺度特征融合的无人机识别与检测

UAV Recognition and Detection Based on Multi-scale Feature Fusion

为了解决复杂背景下的无人机目标识别与检测问题,提出一种改进的更快捷的区域卷积神经网络(fasterregion-based convolutional neural network,Faster R-CNN)模型。针对视野内目标尺寸与标定训练集差异问题,设计了3种不同尺度感受野并行检测、权重共享的特征提取结构——多感受野特征提取网络(trident-visual geometry group network,Tri-VGG),并设计了一种避免网络出现过拟合的训练策略,优化后的网络检测精度(average precision,AP)达到90.9%,实时性达到24帧/秒,满足实际需求。

In order to realize the UAV target recognition and detection in complex background,an improved Faster R-CNN network model is proposed.Aiming at the difference between the target size in the field of view and the labeled training dataset,three different scales of receptive field parallel detection and weight sharing feature extraction structure Tri-VGG network are designed,and a training strategy is designed to avoid network over fitting.The optimized network detection average precision(average precision,AP)reaches 90.9%,and the real-time performance reaches 24 frames per second,which meets actual needs.