A Study on Autonomous Navigation and Obstacle Avoidance Methods for UAVs Based on Visual-Inertial Fusion
This paper proposes a low-cost, lightweight vision-inertial fusion method for micro drones that combines VINS-Mono for high-frequency pose estimation with Ego-Planner for dynamic trajectory planning, achieving robust autonomous navigation and obstacle avoidance in GPS-denied, high-obstacle-density environments with a minimum 70% success rate and superior trajectory smoothness compared to A*.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a small drone trying to find its way through a dark, cluttered tunnel. It cannot rely on the satellite signals that guide our phones or cars, because the rock and earth above block those signals. It also cannot carry heavy, expensive laser scanners, because the drone is too small to lift them. In this difficult situation, the drone must rely on its own eyes and a sense of balance to avoid crashing into walls or obstacles. This is the challenge researchers face when they try to make tiny flying machines operate safely in places like underground mines, disaster zones, or dense forests. For a long time, these machines have struggled to move smoothly or accurately in such tight, signal-free spaces, often leading to crashes or failed missions.
A team of engineers from Taiyuan University and a crane manufacturing company in China has developed a new way for these small drones to navigate. They created a system that combines what the drone sees with how it feels its own movement. The drone uses a single camera to look ahead and a small sensor that measures acceleration and rotation, similar to the one inside a smartphone. By fusing these two streams of information, the drone can figure out exactly where it is and how it is moving, even when it is spinning quickly or the lighting is poor. This system does not need a map of the world beforehand; instead, it builds a picture of the immediate surroundings as it flies, looking for safe paths through the obstacles in real time.
The researchers tested this idea by building a small quadcopter with a wheelbase of 230 millimeters. They equipped it with a camera that takes 30 pictures every second and the motion sensor. The brain of the drone was a compact computer capable of processing data quickly. To see if the system worked, the team set up a challenging test environment. They created a space measuring 4 meters by 4 meters and filled it with 10 to 12 large obstacles, each standing 1.5 meters tall. The lighting was kept very dim, at about 6 lux, which is far darker than a typical living room, and they blocked all satellite signals completely. In this difficult setting, the drone had to fly from a starting point to a target located 5.7 meters away, avoiding the obstacles on its own.
The results showed that the new method was significantly more reliable than older techniques. In these tests, the drone successfully completed its mission 70% to 80% of the time, depending on the specific arrangement of obstacles. When the researchers compared this to a standard navigation method that relies only on visual data without the motion sensor, the older method failed much more often, with a success rate of only 40%. The older method also drifted off course, ending up nearly half a meter away from where it should have been, while the new system stayed within 12 centimeters of the correct path. This difference is crucial because even a small error in a tight space can lead to a collision.
Beyond just reaching the destination, the way the drone moved was also smoother. The new system generated flight paths that curved gently, allowing the drone to glide through the obstacles without sharp turns or sudden stops. In contrast, the older methods often produced jagged, stop-and-go paths that are harder for a drone to fly safely. The researchers found that 87% of the paths created by their new system were smoother than those made by a common planning algorithm. This smoothness is important because it saves energy and reduces the risk of the drone losing control during sharp maneuvers. The system managed to do all this while running on a small computer that did not overheat, even after 15 minutes of continuous flight in a room at 25 degrees Celsius.
The study also looked at how the system handles different conditions. When the light was very low, below 10 lux, the drone's ability to see features in the environment dropped, and the success rate fell to 55%. The researchers noted that in such extreme darkness, the visual sensors struggle to find enough detail to work properly. They suggested that adding a simple infrared light could solve this problem, effectively giving the drone "night vision" to restore its ability to navigate. Similarly, when the obstacles were packed very tightly together, the drone had to work harder to find a path, but it still managed to succeed 70% of the time. The researchers acknowledged that while their system works well in these controlled tests, real-world disaster sites might present even more complex challenges, such as strong air currents or completely dark environments where no light is available.
This work demonstrates that small drones can be made to fly autonomously in difficult, GPS-free environments without needing heavy equipment. By carefully combining visual data with motion sensing and optimizing the software to run efficiently, the team created a lightweight solution that is both accurate and smooth. While the system is not yet perfect for every possible scenario, particularly in total darkness or over very long distances, it represents a significant step forward. The researchers plan to continue improving the technology by adding other types of sensors and testing the system in larger, more complex environments, aiming to bring this capability from the laboratory to real-world applications where it could help in search and rescue or infrastructure inspection.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.