[论文解读] Learning to Navigate Autonomously in Outdoor Environments : MAVNet

Saumya Kumaar Saksena, Arpit Sangotra+4 more|arXiv (Cornell University)|Sep 2, 2018

Robotics and Sensor-Based Localization参考文献 17被引用 6

一句话总结

MAVNet 提出了一种基于模仿学习的端到端导航系统，用于四旋翼无人机，采用一种39层的 Inception-v3 启发式模型，在城市环境中自主追踪道路。该系统基于专家驾驶的无人机影像进行训练，在最高6 m/s的速度下实现了98.44%的道路导航准确率，推理速度达到33 FPS，实现在未训练道路上的357米连续自主飞行。

ABSTRACT

In the modern era of automation and robotics, autonomous vehicles are currently the focus of academic and industrial research. With the ever increasing number of unmanned aerial vehicles getting involved in activities in the civilian and commercial domain, there is an increased need for autonomy in these systems too. Due to guidelines set by the governments regarding the operation ceiling of civil drones, road-tracking based navigation is garnering interest . In an attempt to achieve the above mentioned tasks, we propose an imitation learning based, data-driven solution to UAV autonomy for navigating through city streets by learning to fly by imitating an expert pilot. Derived from the classic image classification algorithms, our classifier has been constructed in the form of a fast 39-layered Inception model, that evaluates the presence of roads using the tomographic reconstructions of the input frames. Based on the Inception-v3 architecture, our system performs better in terms of processing complexity and accuracy than many existing models for imitation learning. The data used for training the system has been captured from the drone, by flying it in and around urban and semi-urban streets, by experts having at least 6-8 years of flying experience. Permissions were taken from required authorities who made sure that minimal risk (to pedestrians) is involved in the data collection process. With the extensive amount of drone data that we collected, we have been able to navigate successfully through roads without crashing or overshooting, with an accuracy of 98.44%. The computational efficiency of MAVNet enables the drone to fly at high speeds of upto 6m/sec. We present the same results in this research and compare them with other state-of-the-art methods of vision and learning based navigation.

研究动机与目标

• 在城市和准城市环境中，实现小型无人驾驶航空器（无人机）的自主道路跟随导航。
• 解决因信号遮挡和精度不足导致的密集城市区域中 GPS 导航的局限性。
• 开发一种计算效率高、可实时运行的导航系统，通过模仿学习模仿专家人类飞行员的操作。
• 通过使用单目视觉和断层扫描重建技术，减少对昂贵传感器（如 LIDAR 或立体相机）的依赖。
• 在不依赖 GPS 或专用传感器的前提下，实现高速、稳定的导航（最高达6 m/s），同时保持道路追踪的高准确率，避免碰撞或越界。

提出的方法

• 基于 Inception-v3 架构设计并训练了一个自定义的39层卷积神经网络 MAVNet，将单目摄像头输入映射为控制指令。
• 使用视频帧（100×100像素）的断层扫描重建作为输入，以增强道路可见性并降低噪声。
• 通过记录专家飞行员在飞行过程中执行的控制动作（前进、左偏航、右偏航、停止、路口）实现模仿学习。
• 模型输出五种离散动作：前进运动、左右偏航、停止，以及用于基于地标导航的路口检测。
• 在包含357米真实城市飞行路径的自定义数据集上进行训练，专家飞行保持在距地2.5米的恒定高度，以最小化对行人的风险。
• 集成了基于简单光流的碰撞避免系统，以应对车辆和树枝等稀疏障碍物。

实验结果

研究问题

• RQ1 轻量级端到端深度学习模型是否能仅使用单目视觉实现实时、高速的城市环境自主道路导航？
• RQ2 模仿学习在复制复杂城市街道中专家人类飞行员行为方面有多有效？
• RQ3 该模型在未训练道路段（尤其是存在路口和光照变化时）的泛化能力如何？
• RQ4 在视觉条件恶劣（如强烈阴影或路面破损）时，模型性能会如何退化？
• RQ5 在不依赖 GPS 或专用传感器的前提下，系统是否能在高速（如6 m/s）下保持高准确率和稳定性？

主要发现

• MAVNet 在未训练的城市环境中实现了357米自主飞行路径的98.44%导航准确率，表现出强大的泛化能力。
• 模型推理速度达到33 FPS（每帧0.03031秒），超过标准30 Hz的摄像头输入速率，支持最高达6 m/s的高速飞行。
• 专家与模型轨迹在位置（X轴）上的相关系数为0.9977，在偏航方向上的相关系数为0.8534，证实了运动控制的高度保真度。
• 系统成功检测到道路路口，并在连续飞行中自主做出转向决策，GPS轨迹图（图11）已验证该结果。
• 在强烈阴影条件下性能下降至81%，原因是边缘伪影，表明对视觉失真较为敏感。
• 模型在具有急转弯和非均匀表面的道路中仍保持稳定与准确，尽管在这些情况下观察到轻微性能下降。