Skip to main content

PercEvite - Sense and avoid technology for small drones

Periodic Reporting for period 2 - PercEvite (PercEvite - Sense and avoid technology for small drones)

Reporting period: 2018-09-01 to 2019-08-31

"PercEvite"" is a project funded by the European organization SESAR, which stands for Single European Sky Air Traffic Management Research. In short, SESAR coordinates developments and research in air traffic management, in order to arrive at common European solutions, shared and supported by all countries.
Air traffic management is going through exciting times due to the booming drone market, as more and more professionally and personally operated drones are taking to the skies. If not managed well, these drones may endanger conventional air traffic (think of rescue helicopters), other drones, and people and property on the ground.

A solution is required that allows drones to detect and stay clear of other aircraft and obstacles on the ground, all by themselves. And it would be most effective if this solution is also suitable for the large group of small drones.
The overall objective of PercEvite is to develop a sensor, communication, and processing suite for small drones. It will enable detecting and avoiding ground-based obstacles and flying air vehicles without necessitating human intervention. To avoid ground-based obstacles, we aim for a lightweight, energy-efficient sensor and processing package that maximizes payload capacity.

The PercEvite consortium consists of two academic and one commercial partner: TU Delft, KU Leuven, and Parrot SA. In the course of three years (starting from September 1, 2017), these partners want to show the potential of a light-weight sense-and-avoid package (in the order of 200 grams), based on frontrunner technologies and extensive real-world tests."
Until now, we have:

1. Developed a multi-communication package with WiFi, LTE, and ADSB-in. This package allows drones to communicate information on their own position to other air users, communicate with other drones, and receive messages from manned aircraft (ADSB-in). Communication will be an important way to ensure sufficient clearance between different air users. For the WiFi, a novel scheme has been developed, which allows much quicker exchange of information between drones – by using the SSID to store position and velocity.

2. Implemented a solution for obstacle detection and avoidance, which uses the two cameras and processing of the Parrot SLAM dunk (a relatively light-weight sensor and processing package). In particular, the method uses both stereo vision (to see distances) and optical flow (to estimate the velocity of the drone). This allows the drone to fly both indoors, where the GPS signal is bad or absent, and outdoors. The drone can fly autonomously and stop if it detects obstacles. We have now ported this system also to a follow-up system with an Intel Realsense stereo vision system and an NVidia TX2 processor.

3. Developed software to see distances in a single image. This way of seeing distances is complementary to using stereo vision (triangulation with two cameras), and has the potential to improve distance estimation. We have further analyzed how this type of network actually sees distances, revealing that it relies on a flat ground plane, and basically determines how high in the image obstacles touch the ground plane. This finding has large implications for the use of such networks on drones (it makes is more difficult).

4. We have created a dataset of helicopter sounds, which can be mixed with drone sounds, and used for the benchmarking the automatic detection of aircraft. We have created a deep neural network that performs the detection based on a spectrogram and a convolutional neural network. The performance of our current network is still not sufficient for real-world use, but the data will be made openly available on the 4TU dataverse, so that the community can improve results also after the end of the project.
In this period we have laid the basis for drones that can avoid each other and other aircraft with a small communication package. The communication package is now complete, with ADSB in, LTE, WiFi, and even a LoRA package. Tests have been performed with drones avoiding each other based on WiFi, and the use of a multi-antenna solution for angle of arrival estimation has been investigated. This last solution, which attempts to use communication from a non-collaborative aircraft currently requires hardware that is too big to fit on small drones. Furthermore, we have implemented a visual navigation system that allows autonomous navigation of our drone, which allows it to fly both indoors and outdoors, stopping in front of obstacles and flying around them. The system won a 1st prize at the IMAV 2018 outdoor competition. Furthermore, we have gained valuable insight into deep neural networks that can see distances in a single image, and into the difficulties involved in the audio-based detection of other aircraft (for hear-and-avoid). The scientific article on hear-and-avoid has won a special achievement award at the IMAV 2019.

The remaining part of the project will focus on creating a really low-weight package (< 50 grams) that allows drones to avoid ground-based obstacles and other air users, combining stereo vision and multiple communication means on a scale that was previously unimaginable. We will also organize a symposium and disseminate the results of our projects (planned for June 2020).