航空学报 > 2025, Vol. 46 Issue (11): 531293-531293   doi: 10.7527/S1000-6893.2025.31293

FLASH:无人机滑坡灾情轻巧感知系统

陈琳, 朱庆(), 胡翰, 丁雨淋, 顾朋鑫   

  1. 西南交通大学 地球科学与工程学院,成都 611756
  • 收稿日期:2024-09-30 修回日期:2024-12-07 接受日期:2025-02-28 出版日期:2025-03-19 发布日期:2025-03-19
  • 通讯作者: 朱庆 E-mail:zhuqing@swjtu.edu.cn
  • 基金资助:
    国家自然科学基金(42230102);国家自然科学基金(42071355);国家自然科学基金(42371436)

FLASH: Flexible and lightweight awareness of slope hazard

Lin CHEN, Qing ZHU(), Han HU, Yulin DING, Pengxin GU   

  1. Faculty of Geosciences and Engineering,Southwest Jiaotong University,Chengdu 611756,China
  • Received:2024-09-30 Revised:2024-12-07 Accepted:2025-02-28 Online:2025-03-19 Published:2025-03-19
  • Contact: Qing ZHU E-mail:zhuqing@swjtu.edu.cn
  • Supported by:
    National Natural Science Foundation of China(42230102)

摘要:

面向断路、断电、断网等极端条件下重大滑坡灾情应急抢险需求,针对当前灾情解译流程时效性差、数据处理链条割裂、灾情解译周期长等问题,设计并研制了一种无人机滑坡灾情轻巧感知系统FLASH。该系统由硬件和算法2部分构成:硬件系统包括轻巧感知和实时解译模块,整个系统设计为可解耦的独立节点,并采用ROS系统消息机制和串口通信实现低时延通信,峰值功率仅25 W,质量500 g;算法部分提出了顾及空间信息约束的轻量化模型,该模型在保持推理精度不下降同时大幅减少超高分辨率影像语义分割的时延,单张7 952×5 304影像推理延迟仅为78.88 ms,并结合深度语义表征实现多影像的滑坡智能检索与方量估算,以松潘滑坡为例,滑坡方量估计可靠性达81%。该系统最终实现了从影像采集到滑坡方量估算的关键灾情信息全流程自动化提取,显著提升了滑坡灾情解译效率。

关键词: 实时滑坡提取, 无人机影像, 应急响应, 边缘设备, 系统集成, 轻量化, ROS系统

Abstract:

To address the urgent needs of emergency response for major landslide disasters under extreme conditions such as disconnected roads, power outages, and network failures, this study designs and develops FLASH, a lightweight UAV-based landslide perception system which integrates hardware and algorithms. The hardware system includes a lightweight perception module and real-time interpretation unit, designed as decoupled standalone nodes. Utilizing ROS-based messaging and serial communication for low-latency data transfer, the system operates at a peak power of 25 W and weighs 500 g. Algorithmically, a lightweight design model incorporating spatial-informational constraints is proposed, which maintains inference accuracy while drastically reducing latency for semantic segmentation of ultra-high-resolution images (e.g., 78.88 ms per 7 952×5 304 image). The model further enables intelligent multi-image landslide retrieval and volume estimation via deep semantic representation, achieving 81% reliability in volume calculations (validated in the Songpan landslide case) without relying on 3D reconstruction. Ultimately, this system realizes the fully automated extraction of critical disaster information from UAV imagery acquisition to landslide volume estimation, significantly enhancing the efficiency of landslide disaster interpretation.

Key words: real-time landslide extraction, UAV imagery, emergency response, edge device, system integration, lightweight design, ROS system

中图分类号: