摘要
在旋翼无人机视频流中,由于道路场景中存在地物遮挡与低对比度光影的干扰,易导致分割歧义问题,故本研究提出基于掩膜区域卷积神经网络(Mask R-CNN)的旋翼无人机视频流实时道路分割算法。针对旋翼无人机拍摄的视频流,考虑到视频流中道路特征呈现多尺度几何异构性与环境噪声耦合干扰,首先构建基于Mask R-CNN的多尺度特征金字塔增强模块,通过参数化伽马(Gamma)分布生成乘性噪声,模拟噪声分布;然后,引入空间自适应注意力窗口,基于像素级真正例率动态调整特征图分辨率,将道路特征值作为拟合函数的参照量,构建参数化函数模型;最后,以实际中道路掩码的几何特征值作为自变量,以视频流中道路目标的实时空间坐标、拓扑结构为因变量,描述几何属性与视频流动态变量间的映射关系,计算道路点云特征相似性,通过点云映射对比实现有效分割。实验结果表明:该方法可有效应对阴暗、复杂和高曝光环境条件的影响,实时道路分割准确性较高,分割后图像的信噪比较高。
In rotary-wing unmanned aerial vehicle(UAV)video streams,road scenes often suffer from object occlusion and low-contrast lighting interference,which can lead to segmentation ambiguity.Therefore,this study proposed a real-time road segmentation algorithm for rotary-wing UAV video streams based on mask region-based convolutional neural network(Mask R-CNN).For video streams captured by rotary-wing UAVs,considering the multi-scale geometric heterogeneity and environmental noise coupling interference of road features,the study first constructed a multi-scale feature pyramid enhance⁃ment module based on Mask R-CNN.Additionally,the noise distribution was simulated by generating multiplicative noise through a parameterized Gamma distribution.Then,a spatially adaptive attention window was introduced to dynamically adjust the feature map resolution based on pixel-level true positive rate.Road feature values were used as reference quantities for the fitting function,and a parameterized function model was constructed.Finally,using the actual geometric feature val⁃ues of the road mask as independent variables and the real-time spatial coordinates and topology of the road targets in the video stream as dependent variables,the mapping relationship was described between geometric properties and dynamic vari⁃ables in the video stream.Road point cloud feature similarity was calculated,and effective segmentation was achieved through point cloud mapping comparison.Experimental data shows that this method can effectively handle the effects of dark,complex,and high-exposure environmental conditions,with high accuracy in real-time road segmentation and a high signal-to-noise ratio in the segmented images.
作者
周玉林
ZHOU Yulin(Shanghai Yingce Surveying and Mapping Services Company Limited,Shanghai 202150,China)
出处
《北京测绘》
2025年第11期1621-1626,共6页
Beijing Surveying and Mapping
关键词
掩膜区域卷积神经网络
旋翼无人机
实时道路分割
道路掩码
伽马分布
mask region-convolutional neural network(Mask R-CNN)
rotary-wing unmanned aerial vehicle(UAV)
realtime road segmentation
road mask
gamma distribution
