融合语义分割与模糊推理的无人机应急降落选址算法  

Integration of Semantic Segmentation and Fuzzy Reasoning for Unmanned Aerial Vehicle Emergency Landing Site Selection Algorithm

作　　者：李迪 肖敏[1,2] 任东[1,2] 谢咏昶 姚远[1,2] LI Di;XIAO Min;REN Dong;XIE Yongchang;YAO Yuan(School of Computer and Information,Three Gorges University,Yichang,Hubei 443002,China;Hubei Farmland Environmental Monitoring Engineering and Technology Research Center,Three Gorges University,Yichang,Hubei 443002,China)

机构地区：[1]三峡大学计算机与信息学院,湖北宜昌443002 [2]三峡大学湖北省农田环境监测工程技术研究中心,湖北宜昌443002

出　　处：《计算机工程与应用》2025年第1期174-185,共12页Computer Engineering and Applications

基　　金：国家重点研发计划(2016YFD0800902);湖北省重大技术创新项目(2017ABA157);湖北省教育厅科技研究计划优秀中青年人才项目(Q20201206);湖北省农田环境监测工程技术研究中心开放基金(201613);111引智项目(D20015)。

摘　　要：随着无人机的应用领域从娱乐摄影拓展到物流、军事和灾害响应,对于无人机的自主智能化要求也越来越高。针对无人机紧急情况下自主降落区域复杂难以保证着陆安全的问题,提出了一种实时语义分割网络与模糊推理相结合的降落选址算法(STDC-LSSNet)。考虑到潜在危险因素在航拍图像上占比小、易被错误分割的问题,提出了小目标特征提取模块(small target feature capture module,STFCM),通过计算不同尺度特征的相似性并进行权重分配,强化小目标特征的表达。考虑到安全区域与危险区域边界混淆会导致无人机降落存在巨大风险,提出了边界特征融合模块(boundary feature fusion module,BFFM),将浅层网络由拉普拉斯卷积得到的边界信息与深层网络的语义信息进行特征融合,引入注意力机制,增强边界区域特征的表达。通过对分割得到的图像进行模糊推理,从而精确识别应急降落地点。所提算法在公开数据集Semantic Drone和AeroScapes上与最先进的算法进行了广泛的对比实验,mIoU提升1.72个百分点和3.89个百分点,实时分割速度达到210 FPS,选址的速度达到58.62 ms,实现了无人机在复杂情况下的应急降落选址。With the expansion of the application fields of drones from recreational photography to logistics,military operations,and disaster response,the demand for the autonomy and intelligence of drones has increased significantly.Addressing the challenge of ensuring autonomous safe landings in complex and unpredictable emergency landing zones,a novel landing site selection algorithm,named STDC-LSSNet(semantic target detection and control for landing site selection net),is proposed by integrating a real-time semantic segmentation network with fuzzy reasoning.Firstly,to address the issue of potential danger factors having a small proportion in aerial images and being prone to missegmentation,a small target feature capture module(STFCM)is introduced.This module calculates the similarity of features at different scales and assigns weights to enhance the representation of small target features effectively.Secondly,considering the risk associated with the confusion between safe and hazardous areas during drone landing due to unclear boundaries,a boundary feature fusion module(BFFM)is introduced.This module combines the boundary information obtained from Laplacian convolution in shallow networks with the semantic information from deep networks,incorporating an attention mechanism to reinforce the representation of boundary region features.Finally,fuzzy reasoning is performed on the segmented images to accurately identify the emergency landing location.The proposed method is extensively evaluated on public datasets,Semantic Drone and AeroScapes,and compared with state-of-the-art algorithms.The results show a significant improvement in mean intersection over union(mIoU)by 1.72 percentage points and 3.89 percentage points.The real-time segmentation speed reaches 210 FPS,and the selection speed achieves 58.62 ms,demonstrating the effectiveness of the proposed approach in enabling drones to autonomously select emergency landing sites in complex scenarios.

关 键 词：无人机 自主降落 实时语义分割 模糊推理 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]