TL;DR
This paper presents a robust system for autonomous indoor navigation of quadrotor helicopters using either stereo vision or laser odometry. By developing a flexible, multi-level sensor processing architecture, the system can leverage the strengths of each sensor independently, paving the way for future integration for even more robust performance in GPS-denied environments.
What Problem Does This Paper Address?
Autonomous navigation for Micro Aerial Vehicles (MAVs), particularly quadrotor helicopters, in indoor environments presents a significant challenge. Unlike ground robots that can rely on GPS or wheel odometry, indoor MAVs operate in GPS-denied spaces and face unique constraints. These include limited payload capacity for sensors and computation, fast and unstable dynamics requiring precise state estimation, and the inability to simply stop and re-evaluate like ground robots. Existing approaches often struggle with these limitations, leading to unstable flight or reliance on external systems. This paper aims to address this by developing on-board sensor-based navigation systems that are robust and adaptable to these challenging conditions.
Methods
The core of the paper lies in developing sensor-specific algorithms for estimating the helicopter's motion and then fusing this information into a unified system architecture.
- Laser Odometry: A high-speed laser scan-matching algorithm is employed. It builds a high-resolution local map from previous scans and aligns incoming scans to this map at 40Hz. This algorithm uses a cost-map derived from polyline contours of the environment to find the optimal rigid body transform, maximizing a reward function.
- Stereo Visual Odometry: A visual odometry algorithm is developed for stereo camera input. It uses a "sparse-stereo" method:
- Features are detected in the left frame and tracked to the right frame for 3D reconstruction via triangulation.
- These 3D features are then tracked from the previous left frame to the current left frame.
- A closed-form solution (Umeyama's method) is used to estimate the relative motion between frames, followed by a robust MSAC refinement.
- To combat error accumulation, a sliding-window bundle adjustment is applied to a window of recent frames, minimizing re-projection errors for improved accuracy.
- System Architecture: A three-level sensing hierarchy is proposed:
- Base Level (1000Hz): Onboard IMU and processor for tight feedback loop stabilization (pitch, roll).
- Real-time Odometry Level: Laser or visual odometry estimates local environment motion. This is fused with IMU data by an Extended Kalman Filter (EKF) for accurate, high-frequency state estimates, enabling stable hovering and basic obstacle avoidance.
- Global SLAM Level (Low Frequency): A SLAM algorithm uses EKF estimates and sensor measurements to build a global map and ensure globally consistent state estimates through loop closures. This module is not part of the real-time loop but provides delayed correction signals.
Datasets
Not discussed in the paper. The experiments were conducted using real-world flights in indoor environments.
Key Insights and Ideas
- Sensor Complementarity: Laser and camera sensors, despite both being exteroceptive, have different strengths and weaknesses. Lasers excel with unique physical structures but struggle with long, featureless corridors. Cameras provide rich 3D information but require distinct visual features. Integrating them offers a path to robust navigation in diverse indoor settings.
- Flexible Architecture: The multi-level sensor processing hierarchy is designed to be sensor-agnostic. This allows for easy swapping of exteroceptive sensors (laser vs. stereo camera) without redesigning the entire system, a crucial step towards integrating multiple sensors.
- Addressing MAV Dynamics: The paper explicitly acknowledges and addresses the unique challenges of MAV dynamics, particularly the need for accurate velocity estimation for stable control, which is often overlooked in traditional SLAM for ground robots.
- Bundle Adjustment for Accuracy: Applying bundle adjustment to a sliding window of frames significantly reduces accumulated error in visual odometry, demonstrating its effectiveness in improving state estimation accuracy over time.
Key Results
- Laser Odometry: EKF state estimates using laser data closely matched ground-truth measurements from a Vicon motion capture system. The average position difference was less than 1.5cm, and the average velocity difference was 0.02m/s with a standard deviation of 0.025m/s over a 1-minute flight.
- Stereo Visual Odometry: EKF state estimates using camera data also closely matched Vicon ground-truth. The paper highlights that bundle adjustment substantially reduced the total error compared to simple frame-to-frame visual odometry.
- Autonomous Flight: The laser-equipped helicopter successfully navigated autonomously in an unstructured indoor environment (MIT's Stata Center) using only on-board sensors, guided by high-level goals clicked on a dynamically built map.
Limitations and Future Work
- Limitations:
- The paper acknowledges that current SLAM algorithms are too computationally intensive for real-time processing on MAVs, necessitating a separate, lower-frequency SLAM module.
- The visual odometry approach relies on feature correspondences, and while robust, it can still be susceptible to outliers and requires sufficient visual features in the environment.
- The laser odometry's 2D assumption might not hold perfectly in highly complex 3D environments, though it worked well in the tested vertical-walled Stata Center.
- Future Work:
- The primary future direction is the integration of both laser and stereo camera sensors onto a single quadrotor helicopter platform. This is facilitated by the developed sensor-agnostic architecture and the availability of a higher-payload helicopter.
- The goal is to leverage the complementary strengths of both sensors to achieve more robust autonomous navigation in fully 3-dimensional, generic indoor environments.
Summary updated