基于模仿学习的光流法单目视觉导航

Imitation learning-based monocular visual navigation using optical flow method

关键词:  无人机  视觉导航  模仿学习  光流  端到端

Key words:  Drone  Visual navigation  Imitation learning  Optical flow  End-to-end

摘要:

传统的视觉导航方案采用感知规划控制三步走的策略，这种模块化的方法局限性非常多，而基于学习的导航方案也面临数据采集费力和泛化能力不足等难题。同时，传统的基于深度信息的方法存在显著弊端：激光雷达体积大、成本高，难以部署于边缘计算设备；双目相机在远距离、低纹理或动态光照场景下的深度估计可靠性较低，且计算资源消耗较大。为解决上述难题，提出了一种基于模仿学习的光流法单目视觉导航方案。整个系统分为3个部分：拥有全局信息的专家模块、配备光流网络的学生策略和一套训练管理流程。专家模块通过Metropolis-Hastings（M-H）采样计算局部无碰撞轨迹，并且使用改进的A*算法得到全局无碰撞轨迹，作为M-H采样的初值。学生策略创新性地将一个轻量级光流网络融合到避障深度神经网络中，该网络具有图像、状态、融合及规划等分支，其中图像分支使用MobileNet-V3提取图像特征。训练管理流程在基于Unity3D和Flightmare仿真器的仿真环境中，使用数据集增强采集（DAgger）算法进行训练和测试，为专家和学生提供交互界面。实验结果表明，相较于传统的深度感知方案，光流法在一定速度区间内的表现更佳，在4~6 m/s范围内的成功率超过80%，且由于使用的是单目RGB相机和轻量级光流网络，整个系统的成本更低，计算资源消耗也更少。

Abstract:

Traditional visual navigation approaches follow a three-step strategy of perception-planning-control, which exhibits multiple limitations due to its modular nature. Learning-based navigation solutions also present challenges, including the need for laborious data collection and insufficient generalization capability. Meanwhile, conventional depth-based methods have significant drawbacks: LiDAR systems suffer from bulky size and high costs, hindering edge deployment, while stereo cameras exhibit low reliability in depth estimation under distant, low-texture, or dynamic lighting conditions, and require substantial computation. To address these challenges, an imitation learning-based monocular visual navigation scheme using optical flow is proposed. This system comprises three components: an expert module with global information, a student policy equipped with optical flow networks, and a training management pipeline. The expert module calculates local collision-free trajectories through Metropolis-Hastings (M-H) sampling, and uses an improved A* algorithm to generate initial global collision-free trajectories, which provide an initial value for the M-H sampling process. The student policy innovatively integrates a lightweight optical flow network into an obstacle-avoidance deep neural network, featuring branches for image processing, state analysis, feature fusion, and planning, among which the image branch uses MobileNet-V3 for feature extraction. The training pipeline implements the dataset aggregation (DAgger) algorithm within a Unity3D and Flightmare simulation environment to facilitate data collection, training, and testing, and also provides interactive interfaces for expert-student collaboration. Experimental results demonstrate that compared with traditional depth perception-based approaches, the optical flow method achieves better performance within specific velocity ranges, and attains success rates of over 80% in the 4~6 m/s range. Furthermore, the system reduces costs and computational resource consumption through its monocular RGB camera configuration and lightweight optical flow network architecture.