- This paper introduces Edge-FS, a computer vision algorithm enabling efficient velocity estimation and obstacle avoidance for autonomous pocket drones.
- Experiments validated the system, demonstrating sustained autonomous obstacle avoidance flights exceeding 100 seconds in cluttered environments.
- The research significantly advances MAV capabilities, enabling autonomous operation in GPS-denied environments with minimal external hardware requirements.
Efficient Optical Flow and Stereo Vision for Velocity Estimation and Obstacle Avoidance on an Autonomous Pocket Drone
The paper presents a computer vision algorithm, Edge-FS, designed for efficient velocity estimation and stereo vision on autonomous Micro Aerial Vehicles (MAVs). This algorithm allows small pocket drones, weighing less than 50 grams, to autonomously navigate without relying on GPS or significant external computing resources, a marked challenge due to these platforms' computational and power constraints.
Technical Contributions
Edge-FS leverages a lightweight stereo camera setup with an embedded STM32F4 microprocessor operating at 168 MHz. The novel approach combines refined versions of EdgeFlow and EdgeStereo algorithms to deliver both velocity and depth estimations necessary for stabilizing inherently unstable flying platforms, such as quadcopters, without burdensome computational demands. Operating at 20 Hz, the approach employs edge distributions to compute optical flow and stereo disparity, crucial for determining the drone's velocity and avoiding obstacles during flight.
Comparative Evaluation
One of the key technical evaluations carried out in this paper is the comparison of Edge-FS against the Färneback dense optical flow method—a standard optical flow technique also used in robotics applications. The evaluation demonstrated that Edge-FS offers more accurate velocity estimates with lower Mean Squared Error (MSE) and higher Normalized Maximum Cross-Correlation Magnitude (NMXM), albeit on low-resolution images (128 x 96 pixels). The lightweight nature of the algorithm ensures that it situates well within the processing limits of pocket drones without compromising performance or autonomy.
Experimental Validation
Experimental results validate the practical applicability of Edge-FS for autonomous obstacle avoidance in constrained environments. In test scenarios, pocket drones equipped with the stereo vision system were able to sustain autonomous flights lasting beyond 100 seconds in cluttered environments. The system's performance was robust, albeit subjected to certain limitations in depth measurement accuracy at long distances due to the small camera baseline.
Implications and Future Directions
The paper's contributions have significant implications for the development of MAVs. By achieving meaningful velocity and obstacle avoidance with minimal hardware, Edge-FS extends the operational capabilities of MAVs in GPS-denied environments. These technological advancements minimize reliance on external systems, making MAV operations more efficient and independent.
It is worth considering future directions where further tuning of obstacle avoidance logic could enhance MAV navigation reliability. For instance, integrating additional lightweight sensors or improving the field of view could address challenges in environment sensing. This could further expand the range and complexity of tasks that MAVs can autonomously execute.
In conclusion, this paper underscores the potential to leverage efficient computer vision algorithms for extending the capabilities of micro aerial robotics. Edge-FS offers a viable solution to the fundamental challenge of autonomous MAV deployment in indoor environments, opening avenues for enhanced autonomy in diverse applications ranging from search and rescue to precision agriculture.