Drone delivery detection dataset
Matrix audio was recorded at 16Khz. Onboard audio was recorded at 48Khz and resampled at 16Khz. The drone.wav is the audio of reference and it's starting time is saved in the StartTimestamp.csv file.
- drone.wav -> Contains the onboard audio.
- matrix.wav -> Contains the audio from the microphone array. The microphone array is synced with the onboard audio. This is why some recording contains silence at the beginning. Also, some flight weren't recorded with the microphone array so this file might be missing in some folders
Every csv file contains the data columns and a timestamp in Unix Timestamps. The timestamp column indicate the time at which the data point was received by the Raspberry Pi. The seperator of the csv is ; and it can contains N/A values
- acceleration.csv -> Contains the acceleration of the drone documentation
- actuator.csv -> Contains the PWM signal sent to the 4 motors. Some flights directly have the BPF value in the actuator file instead of the raw pwm value documentation
- isflying.csv -> Contains a flag that indicates if the drone is flying or not (1 = flying, 0 = not flying) documentation
- orientation.csv -> Contains the orientation (pitch, roll, yaw) documentation
- velocity.csv -> Contains the speed of the drone (Vx, Vy, Vz) documemtation
- position.csv -> Contains the GPS position of the drone documentation
- pwm.csv -> Contains the PWM signal sent ot the drop mecanism servo motor. A value of 1900 means that the servo is closed and a value of around 1300 means that the servo is open. The PWM value are approximative and should be used with caution. We recommand to look for value between 1100 and 1400 for delivery moment. There are also moments where the servo is open and no delivery was made. To be sure that a given moment has a delivery, it is recommended to also look at the motor speed. (Some flight weren't recorded with the delivery mecanism so this file might be missing in some folders)
- StartTimestamps.csv -> Contains the drone.wav starting time in Unix Timestamp (ms). The Timestamps doesn't necesseraly match the date of the flight since the Rpi wasn't connected to the internet. Use this timestamps to synchronize data in the csv files to the audio files.
- additional_data.json -> This files contains general information like microphone array position and microphone index to ignore. In some case, single microphone picked up electromagnetic noise, making the channel useless. We removed them and
This folder contains all the deliveries extracted from the different flights. Data collection were done at different time and contains different data. Some folders contain the microphone array audio and the onbaord audio while others do not have the array audio. Some also contains the ground truth BPF(actuator.csv) but each contains the BPF predicted by the newtork (drone_predictions.csv and matrix_predictions,csv). Like mentionned in the article, the delivery are not always perfectly synced with the real delivery moment .i.e when the motor speed drops. We suggest to either offset the delivery moment or to consider a window besides the delivery moment as valid. Note that for this folders, the actuactor in the actuactor.csv file contains the raw pwm value.
This folder contains the drone samples and the background noise samples that were used for the SNR performance evaluation. Those samples were extracted from other flights so only the minimal data is present.
This folder contains the flights made for evaluating the network performance at different distances.
Those folders contains the data used for training the neural network. Data augmentation was done at runtime so only the clean data is present.
Noise sample that was manually selected and added to the training dataset. This small datatset contains aircreaft, helicopter and animal noises. All the noises were downloaded from Pixabay
This directory contains all the other flights that weren't included in the other folders. They are the first flight that were made and most are missing some data. For those flight, the StartTimestamp of the drone.wav file was not saved automaticly meaning that there is an offset between the first telemtry data points and the first audio point. Also, since no microphone array was used, the reference position is simply the first GPS point of the drone.
The GPS position of the drone are saved in most of the flight and the microphone array position is also saved. However, the microphone array was not aligned with some other geographic marker like the true Notyh. For example, lets say you localize the sound source using the array audio and compare the results to the GPS position, you WILL observe some offset in the data in the azimuth and elevation angles. Moreover, the alitutde of the drone is less reliable than the Latitude and Longitude data. We observe some weird jump in the altitude mid flight. The GPS position for the flight should only be used as an overview or to compute an approximative distance from the microphone array GPS position.
If you want to perform localization on the matrix audio, please use a small localization sample and the corresponding true gps position to align them.
For questions, please reach out to [email protected]