融合HC-FPN与ICRU的机载视觉滑行道线检测方法

A Taxiway Line Detection Method for Onboard Vision Based on the Fusion of HC-FPN and ICRU

摘要:

为通过机载视觉传感器获取飞机滑行方向与滑行道线的相对位置关系，增强民航飞机滑行阶段机时机组人员对滑行道线的感知能力，以实现民航飞机视觉辅助驾驶，提出了一种融合HC-FPN与ICRU的机载视觉滑行道线检测方法；通过ResNet进行图像特征提取，在网络颈部设计了HC-FPN融合多级高层信息；通过实例检测，并采用行锚检测与动态卷积预测滑行道线形状；此外设计了ICRU满足具有复杂拓扑结构的滑行道线检测；在损失中引入了PIoU损失函数；经过实验测试，该网络在自制滑行道线数据集上F1达到85.81%，在CurveLanes数据集和CULane数据集上F1达到86.17%和78.63%，在TuSimple数据集上准确率达到96.45%，实验结果证实了该方法的准确率和有效性，有助于飞机滑行过程中机组人员辅助驾驶。

Abstract:

To acquire the relative positional relationship between the taxiing direction of an aircraft and taxiway markings using airborne vision sensors, enhance the crew's perception of taxiway markings during the taxiing phase of civil aviation aircraft, and achieve visual-assisted driving for such aircraft, an airborne vision taxiway line detection method integrating HC-FPN and ICRU is proposed. Image features are extracted using ResNet, and HC-FPN is designed in the network neck to fuse multi-level high-level information. Instance detection is performed, employing row-anchor detection and dynamic convolution to predict the shapes of taxiway lines. Additionally, ICRU is designed to meet the requirements for detecting taxiway lines with complex topological structures. The PIoU loss function is introduced into the loss calculation. Experimental results show that the proposed network achieves an F1 score of 85.81% on a self-developed taxiway line dataset, 86.17% on the CurveLanes dataset, 78.63% on the CULane dataset, and an accuracy of 96.45% on the TuSimple dataset. These results confirm the accuracy and effectiveness of the proposed method, which aids crew members during aircraft taxiing for assisted driving.