A vision-based autonomous UAV inspection framework for unknown tunnel construction sites with dynamic obstacles

TL;DR

The paper addresses safe autonomous inspection of drill-and-blast tunnel fronts in unknown, dynamic environments without prior maps by proposing a vision-based UAV system with RGB-D sensing. It introduces a hierarchical planning framework coupled with a 3D dynamic map that separately represents static obstacles via a voxel grid and dynamic obstacles via bounding boxes, enabling fast, gradient-based obstacle avoidance at run time. Key contributions include a high-level task planner (Forward/Explore/Inspect/Return), a perception module that fuses depth data with VIO via EKF, a dynamic map with Kalman-tracked obstacles, and a low-level trajectory optimizer using a receding horizon distance field for dynamic safety. The approach is validated through simulations and real tunnel tests, achieving safe navigation, effective dynamic obstacle handling, and accurate 3D reconstruction with mean errors on the order of several centimeters, highlighting practical potential for construction-site safety and efficiency.

Abstract

Tunnel construction using the drill-and-blast method requires the 3D measurement of the excavation front to evaluate underbreak locations. Considering the inspection and measurement task's safety, cost, and efficiency, deploying lightweight autonomous robots, such as unmanned aerial vehicles (UAV), becomes more necessary and popular. Most of the previous works use a prior map for inspection viewpoint determination and do not consider dynamic obstacles. To maximally increase the level of autonomy, this paper proposes a vision-based UAV inspection framework for dynamic tunnel environments without using a prior map. Our approach utilizes a hierarchical planning scheme, decomposing the inspection problem into different levels. The high-level decision maker first determines the task for the robot and generates the target point. Then, the mid-level path planner finds the waypoint path and optimizes the collision-free static trajectory. Finally, the static trajectory will be fed into the low-level local planner to avoid dynamic obstacles and navigate to the target point. Besides, our framework contains a novel dynamic map module that can simultaneously track dynamic obstacles and represent static obstacles based on an RGB-D camera. After inspection, the Structure-from-Motion (SfM) pipeline is applied to generate the 3D shape of the target. To our best knowledge, this is the first time autonomous inspection has been realized in unknown and dynamic tunnel environments. Our flight experiments in a real tunnel prove that our method can autonomously inspect the tunnel excavation front surface. Our software is available on GitHub as an open-source ROS package.

Figures (7)

• Figure 1: Illustration of UAV navigating and inspecting the excavation front in the tunnel environment. (a) The tunnel under construction. (b) The target inspection area (the excavation front). (c) The robot navigates toward the inspection target and avoids obstacles. (d) The robot inspects the target area.
• Figure 2: System framework for autonomous inspection. Our proposed framework contains three parts: visual perception, hierarchical planning, and data post-processing. In the visual perception step, the localization module applies the visual-inertial odometry with EKF fusion for state estimation. The dynamic map module builds the static voxel map and tracks dynamic obstacles based on depth images. In the hierarchical planning section, the high-level and mid-level planners use the static voxel map to generate the static trajectory. Then, the low-level planner uses the dynamic obstacle information to optimize the output trajectory for execution. The final data post-processing step takes the images collected from the inspection stage to reconstruct the target model for analysis.
• Figure 3: Illustration of the proposed 3D dynamic map. (a) A person walks in front of the robot in the RGB camera view. (b) The person is detected as a dynamic obstacle in the depth image. (c) The detection results in the U-depth map for obstacle widths and thicknesses. (d) The 3D dynamic map shows the dynamic obstacle as a bounding box and static obstacles as the voxel map.
• Figure 4: Illustration of the collision cost in our B-spline optimization. (a) The static collision cost is calculated using the proposed circle-based guide points (red dots). (b) The dynamic collision cost is obtained by the receding horizon distance field, which considers the future predictions of the obstacle positions.
• Figure 5: Illustration of an example simulation tunnel environment in Gazebo. In the forward task, the robot needs to navigate from the tunnel start (left side) to the tunnel end (right side) and avoid static and dynamic obstacles.