Recently, unmanned aerial vehicles (UAVs), or drones, have increasingly been found in many real-life applications including search and rescue, warehousing, agricultural monitoring, and military-grade missions. In such applications, accurate navigation, control and keeping the drones safe while operating in dynamic and unknown environments remains one of the top priorities. Thus when wind gust arises in the flight path of the drone, we look to implement a cost effective and non-bloated method of detecting the speed and direction of the wind gust. Adding an anemometer will cost more and add weight, especially to Micro Air Vehicles (MAV, UAV) that are already weight and payload limited. In this work, we discuss and report the feasibility of using commodity physical sensors, e.g., accelerometers and gyroscopes, to help drones sense wind gusts.
Accelerometers and gyroscopes have been used successfully in detecting human activities [1, 2]. In that line of work, such sensors are deployed in daily used devices such as smart phones , smart watches [4, 2], or health devices, e.g., Fitbit . These works showed that as human users walk, eat, or jog, their body, hand, and arm movements will trigger such sensors to change their readings in manners that allowed highly accurate recognition of such activities.
Inspired by the success of such work, we propose to apply similar mechanisms to equip drones with the ability to detect wind gust. The intuition is that a key environment factor, e.g., air stillness, will most likely be disturbed by anything that comes with motion, such as wind. If the drone is able to accurately associate the existence of wind with the change in its sensory readings, it can track and infer whether there is any such entity around.
The organization of the paper is as follows. We first survey related works in detection for drones. Next detail different components of the environment sensing task, including it being cast as a classification problem and how it can be incorporated into a general wind-detection framework. We then discuss our empirical study in assessing how feasible this approach in real-life situations. The paper concludes with discussions on existing limitations of our study, as well as implications for future works.
Ii Related Works
Environment sensing refers to the task of detecting other entities populating the same physical space in which the drone is operating. These entities can come with a physical body, such as walls, obstacles or other robots, or none at all, such as wind gusts, air turbulence, gas, or interfering radio-frequency signals.
Detecting visible entities have been the main subject of study in most research on environment sensing, the task of which is often achieved using vision-based approaches. For instance, Engel et al.  used a monocular SLAM system to detect objects in a scene, allowing the drone to navigate through unknown environments, without the need of artificial markers or external sensors, in areas where GPS cannot be accessed. Similarly, Ross et al.  demonstrated how on-board monocular vision can be utilized to assist in the navigation and guidance of micro aerial drones through cluttered environments at low altitude, such as outdoor forests. Many of these achieve this through feature detection with SIFT/SURF(Speeded Up Robust Features) methods [8, 9]. In this line of work, applications of environment sensing often include localization of the drones , or for object detection and avoidance .
In contrast, our work aims to detect hazards that do not take a physical or visual form, such as wind gust, or even when obstacles are out of the field of view of the cameras on the drone.
In detecting invisible entities such as gas leakages, a popular approach is the use of specialized sensors attached to their drones, such as micro-machined gas sensors for the purpose of monitoring and detection of gas leakage localization [12, 13]. While these approaches work for special environment entities such as gas, it is unclear whether a sensor can be designed to reliably detect wind in real time especially for use on drones, as the propellers themselves may alter the airflow around the drone where the sensor would be located. Notably, most works in detecting wind gusts focus on the negation of such wind effects in controlling the drone, to create a stable flight [14, 15]. Our work focuses more on detecting aspects of wind gusts that can be picked up by the drone’s sensors, thus opening doors for other downstream tasks.
Environment sensing may be conducted on a broader scale than just a vicinity surrounding a drone. For instance, some works used UAVs equipped with cameras for remote sensing and monitoring of forestry and coastal environments [16, 17]. Such approaches yield many important and practical applications, such as surveying forests, mapping canopy gaps and canopy height, and tracking forest wildfires .
Finally a paper from the Defense Science and Technology Group on Low Airspeed Measuring Devices for Helicopter 
, though lacking in technical detail due to military related work. It did a great job in shading light on interesting methods in detecting low airspeed were mechanical (mechanical anemometer) and non-mechanical system (These include GPS and laser based systems, algorithm systems and neural networks).
Iii Physical Sensors in Drones
In this work, we used CrazyFlie 2.0, a product by BitCraze , for our experimentation. This MAV comes with the following on-board sensors:
3 axis gyro (MPU-9250),
3 axis accelerometer (MPU-9250),
3 axis magnetometer (MPU-9250),
To increase the range of sensors available, it is often possible to augment the CrazyFlie 2.0 drones with additional sensors. Some sample specialized decks, which can be mounted on a vanilla CrazyFlie, include Flow deck, LED-ring deck, or Locopositioning deck, which can all be mounted down-facing on the CrazyFlie drone. Flow decks can enhance hovering and flight stability, using optical flow and Time-of-Flight (ToF) laser-ranging module. LED-ring deck is an array of 12 RGB LEDs that can be used to create light shows when grouped with other drones. Finally, the Locoposition deck can be used as part of an indoor localization system, which sends short high frequency radio messages between the Anchors and Tag (mounted on the Locoposition deck on the drone) to measure, through the means of triangulation, the absolute position of the drone.
For our work, Flow decks was mounted to the base our drones to ensure the stability of flight. This is because the current drivers that accompany the drones require this deck to maintain flight path and/or position.
Iv Detection Framework
The physical sensor data as described in Section III are used to keep the drone stable and provide further information on its physical states. By utilizing such information for the purpose of environment sensing, we propose to push the limit of such data to achieve real-time wind gust detection. Fig 1 depicts the basic components of such a framework.
Our framework first trains detection models during an offline phase and then applies such models in the real-time phase. In both offline training and real-time phases, further data pre-processing using Fast Fourier Transform (FFT) is applied in the hopes of boosting model performance.
Iv-a Building a Wind Gust Detection Model
Our framework is composed of two phases, an offline training phase (split into wind speed and direction detection) and an online real-time tracking phase. In the offline training phase, an environment sensing model is built to associate the physical sensory data as input with corresponding environment factors that affect the drones’ stability. For instance, the model can capture the relationship between wind direction and strength, from the gyroscope and accelerometer readings. This task is similar to the works in activity recognition , in which the goal is to associate human activities with such physical data. While for the online phase, the data is actively collected and analyzed immediately with serialized models from the offline phase.
In general, classification and regression models can be employed to achieve this task, depending on the nature of the environmental factors to be recognized. For example, wind strengths/speed can be categorized or regressed from input data, while wind direction may be categorized into different discrete directions and classified accordingly.
Furthermore, in real time wind gust detection, we look to use Fast Fourier transform (FFT) to help improve detection accuracy. In our study, FFT acts as a filter, as it removes the low frequencies that are unimportant given the effects of wind gust. We also propose a generalized method for applying FFT to our type of data by picking the most dominant “k” frequency domains from a sample of data and establishing out cutoff point for FFT with these frequency domains.
V Empirical Study
To validate the feasibility of environment sensing using drones’ physical sensors, we conducted an empirical study, in which the goal is to predict wind strengths given sensory data collected from the accelerometer and gyroscope. In the first study, we look at the feasibility of wind gust detection in terms of wind speed, wind direction and finally realtime analysis with pre-trained models while the drone is hovering. Then in study 2, we move to further testing of wind direction for drones in-motion.
V-a Data Collection
We used CrazyFlie 2.0 drones as our target drones to collect data from. BitCraze’s CrazyFlie 2.0  is a family of palm-sized micro-drones that was designed for lightweight and versatile maneuver.
The data used in this study were collected from two main sensors, the accelerometer and gyroscope. A third set of data was collected from stabilization values the drones outputted. The stabilizer values are also generated by the CrazyFlie drones which is the fusion of raw gyroscope and accelerometer data using an open-source algorithm called AHRS. Readings from the magnetometer (measuring orientation) and barometer (measuring pressure) are not studied as they are irrelevant to the stability of the drone when subject to wind gusts.
For Study 1(Section VI), the entire data set is collected through multiple flights due to the limited capacity of the battery. Since we will transform the data using a 1-second overlapping window (as discussed later in Section V-A3), total flight time for each packet was set at 64 seconds and we discarded 2 seconds of data points at the beginning and end of the each packet to make sure that no takeoff and landing turbulence may affect our data readings. This yields 6000 data points each flight data packet. For each wind speed or direction, a total of 6 packets were collected. 4 drones were used for wind speed and 3 were used for wind direction. Giving us a total of almost a million usable data points.
To ensure consistency in data from one reading to the next, we made sure the drone stayed within 10 cm of its takeoff mark and discarded readings when the drone’s flight deviated outside of this area. Total flight time for each packet was set at 64 seconds and we also discarded 2 seconds of data points at the beginning and end of the each packet to make sure that no takeoff and landing turbulence may affect our data readings.
Similarly for Study 2 (Section VII), data was collected over multiple flights, however this time it was not only due to battery constraints, but also length of flight zone. Again we will transform the data using a 1-second overlapping time-window, and total flight time is about half a minute. Generating just over 2000 data points per data packet for each flight. Also to make up for the reduced data point from each flight, 10 packets were collected for each set of label instead of 6. 3 drones were used to collect data, but 1 drone’s dataset was corrupted, thus we had to discard it. However we still had over 200,000 usable data points from the 2 drones for data analysis in this study.
To create air turbulence, we used a box fan, set up to blow air right into the drone. The setup for study 1 of the data collection is depicted in Fig 2. We collect sensory data from a hovering drone in four different conditions: No wind, Wind speed 1m/s (meters per second), Wind speed 2m/s and Wind speed 3m/s. The wind speeds are measured using a hand-held anemometer. The drone was also set to fly in a zone that has a consistent desired wind speed. For wind direction, the wind speed was set at 1.5 m/s, then the drone rotated in each direction to record head wind, tail wind, wind from left and right side. The head/front of the drone was determined by marking on the Crazyflie 2.0 that states the front of the drone.
Since we only wanted readings from the drone once it was hovering, we excluded data during take off and landing. The hovering height of the drones was set to 0.4 meters above the ground. This height value is selected such that the drone’s hovering position lies right in the center of the air current produced by the fan. The drone was also set to fly at 0.5 meters away from the fan to ensure that the generated airflow covers a substantial area around the drone, and stays consistent throughout the data recording phase.
For study 2, depicted in Fig 3. We collect sensory data from a drone in motion. It moves from a set point A (1m/s) to B (2m/s) marked on the floor, which we will call the flight zone. Thus for front/head wind, the drone will fly from point A to point B and back/tail wind, the drone will fly in the opposite direction (i.e point B to point A). For side wind, the box fan is set on the marked line at the bottom of the Fig 3, and it moves alone the with the drone as it flies from point A to B for left wind and vice versa for right wind.
This flight zone is limited between 1m/s and 2m/s wind speed when flying either against or away from the wind, because we wanted consistent data for each packet we collected. Thus if speed drops below 1m/s we feared the drone may not pickup the wind gust and if it goes above 2 m/s there is increase of turbulence which will throw the drone off course and render the packet of data useless.
V-A2 Data Description
The tables below show a snippet of the data we collected from the drones. The data are all time-stamped and for gyroscope (Table I) and accelerometer (Table II) the three axes are x, y, and z. While for stabilizer (Table III) the three axes are roll, pitch and yaw.
V-A3 Feature Engineering
Since conventional machine learning models do not work well with time-series data, we pre-processed the data by segmenting it into examples using a sliding window of 1 second. Since the data is collected at 100Hz, one data window comprises of 100 raw data. Next, each data window from each sensor, i.e., accelerometer, gyroscope, and stabilizer, were used to extract 40 high-level features, including:
Mean (3): Mean sensor value (each axis).
Standard deviation (3): Standard deviation (each axis).
Average Absolute Difference (3): Average absolute difference between the 100 values and the mean of these values (each axis).
Average Resultant Acceleration (1): For each of the 100 sensor values in the window, take the square root of x, y, and z axis values and then average them.
Binned Distribution (30): 10 equal-sized bins are formed using the range of (maximum - minimum) of the 100 values, and record the fraction of the 100 values within each bin.
The reason for choosing such features is that, we look to preserve the trend of the data series with these simple hand crafted feature. Also in discussion with Dr. Weiss, we settle on these simple feature as more sophisticated mathematical features such as Fourier analysis did not yield much better results than these. In the end of feature transformation, we are able to keep an overall data trend while defining a more prominent averaging trend line.
Vi Study 1: Wind Gust Detection for Hovering Drone
In our first study, our goal is to examine the feasibility of detecting wind speeds and direction of the wind, i.e., the training and test data for a model come from the same drone, then the trained model is tested using test data from other drones to check generalization. Finally we look to test the real-time capabilities of our models.
Vi-a Model training
We compare the performance of standard classification algorithms in detecting the wind condition under which observed data was recorded: Gradient Boosting (GB). We selected this models due to their superiority in modeling non-linear relationships in data.
We split the entire data set into training test sets, with ratio 4:1, i.e., 80% training and 20% testing. We generate data features for the training and testing data separately. For each model, we applied grid search algorithm with 3-way hold out cross-validation to tune hyper-parameters on the training set. Finally, the left-out test set was used to evaluate each model. As for input features, we tested with data from accelerometer and gyroscope. In our study, we tested both classifier and regressor. For regressor, we measure model performance with r-squared value on a scale of 0 to 1, with any negative value representing an arbitrary bad model.
Vi-B1 Wind Speed
In wind speed detection, we achieved exceptional results using gyroscope data which can be seen in Table IV. Gyroscope Data using training and testing data from the same drone (within drone) yielded accuracies in the range of 85.9% to 92.7%. Even when using training data from one drone and testing with data from other drones (across drone), seen in Table V
we observe a high accuracy in the range of 80.2% to 94.7%. This proves the generalizability of the model for wind speed detection, in that we may use the same model across drones of the same class. In most cases when accuracy’s are this high there isn’t much to infer from the confusion matrix, however, if we observe the within drone and across drone confusion matrix TablesVI and VII, we notice a trend in the misclassifications. The number of misclassifications increase as our wind speed increases, this we believe, is due to the increase in turbulence at higher levels of wind speed, thus altering the data.
A reason why we did not test wind speed above 3 m/s is also attributed to the increase in turbulence at higher wind speed. At above 3 m/s, the drone would have a hard time taking off and maintaining it’s flight position, thus making it impossible to collect meaningful consistent data for analysis. Also the reason for not breaking down the wind speeds into smaller increments is because in a previous experiment, we used much smaller increments of 0.4 m/s and noticed that there was quite a lot of misclassifications, especially across drone testing. This was because the change in data was not significant enough for the model to classify from one dataset to the other.
|Gyroscope Accuracy (%)||92.7||90.7||87.7||85.9|
|Accelerometer Accuracy (%)||71.3||79.5||53.7||72.9|
|Train/Test||Drone 1||Drone 2||Drone 3||Drone 4|
|Drone 1 (%)||-||91.8||87.4||87.4|
|Drone 2 (%)||91.4||-||84.0||85.0|
|Drone 3 (%)||94.7||91.5||-||86.1|
|Drone 4 (%)||89.2||87.0||80.2||-|
|0 m/s||1 m/s||2 m/s||3 m/s|
|0 m/s||1 m/s||2 m/s||3 m/s|
Vi-B2 Wind Direction
In wind direction detection, the accelerometer accuracy was much higher, all above 95%. However, gyroscope data left much to desired, being as low as 31.3%, which can be seen in Table VIII. However, when testing data across drone for accelerometer data, we see a large variation in accuracy ranging from as low ass 51.8% to 80.5%. Thus we would like to look at the confusion matrix to see if we can find a trend with why the classifier performed poorly in across drone analysis. However when looking at the confusion matrices for the across drone analysis, Table X show most of the misclassifications in back wind classified as front wind, while in Table XI the misclassifications are in completely different and random areas.
Since the drones in this study are hovering, there should be no resultant acceleration. Thus accelerometer data should have had no influence on the outcome. Furthermore, the high accuracy of accelerometer data for within drone classification can be attributed to over-fitting of the classifier to the minute changes in data when adjusting to wind from different direction and also shown in the high training data accuracy, while having much lower test accuracy when testing with different data.
This led us to move forward in our data collection, by collecting directional data for drones in-motion which we dive into in study 2.
|Gyroscope Accuracy (%)||45.1||54.7||31.3|
|Accelerometer Accuracy (%)||95.8||98.2||95.6|
Vi-B3 Real-time Analysis
For real-time analysis we built a system that collects the data from the drones and outputs a prediction in real time. For our experimental purposes, the setup itself is similar to that of the offline phase. We then marked the different wind speeds on the ground shown in Fig 2 and changed the location of the fan accordingly. Once the drone is hovering, we collect data in small batches of a 100 data points (rows). We then apply FFT and feature generation on the data. Finally, we test the data using saved serialized models from the offline training phase and record the outputs over a one-minute interval. This is done to allow us to draw accuracy reading from the testing. However, to fully test the real-time capability of our models, we vary the wind speed actively and seeing how the model responded. A video footage of our real-time detection experiment is provided on YouTube111https://youtu.be/hZRcSMr0DFE.
We were able to achieve an output of result at an average rate of one every 0.04 seconds (i.e. 25Hz). Real-time testing proved the strength of our wind speed detection model, where we were able to achieve 100% accuracy when no wind was present, 92%, 86% and 98% accuracy for wind speeds 1m/s, 2m/s and 3 m/s respectively.
However, the accuracy of wind direction detection in real time was no better than guessing when testing. The randomness in prediction when varying the direction of wind acting on the drone meant we weren’t able to collect any meaningful data for analysis.
Vii Study 2: Wind Direction Detection for In-motion Drone
In our second study, our goal is to examine the feasibility of detecting wind direction of the drones when in-motion. This study aims to extend the analysis from study 1, specifically the directional analysis model.
Vii-a Model Testing
We employ similar method of testing the data from study 1. But with stabilizer data. This is due to observations made in study 1, were hovering drone meant that there was no resultant acceleration. Thus with the introduction of motion into the analysis, we must include both gyroscope and accelerometer data, and since Stabilizer data is a fuse of the 2 as mentioned above, it naturally became the best data type for this study.
For the 2 drones used in study 2, we received an accuracy of 62.36% and 57.81% for drone 1 and 2 respectively. This is much lower than the accuracy we desired, which should be around 90% to allow accurate response of the drones to the predictions. However, when looking at confusion matrices in Tables XII and XIII for both drones, we notice the diagonal line of correct classification still contains the majority of the data points, but there are quite a lot of misclassifications throughout and concentrated in the left and right wind area. So the question becomes, why the model failed for this data, but also what was the model classifying. We will break it down to 2 methodologies of explainging this. 1, Data Similarity and 2, Power Spectral Density.
Vii-B1 Data Similarity
Thus to test mathematically if there is any opposing trend between the different directions of wind (i.e. left vs right, front vs back) to explore 3 method of calculating the similarity in data. These being Euclidean Distance, Dynamic Time Warping and Cross Correlation. For Euclidean Distance and Dynamic Time Warping because there is no meaningful benchmark for which we can compare the opposing datasets, we thus decide to use Cross Correlation. With Cross Correlation, correlation at lag of 0 (i.e. X lags mean we shift a sequence to the left or right by x displacement.), data is normalized by the correlation model and resulting values fall between -1 and 1 for easy interpretation. Were -1 is negative similarity and 1 is strong positive similarity.
However when viewing the resulting correlation scores for the different directions in Table XIV, we notice that all score hover around 0, which means that there is no similarity or correlation between the different directional data, thus there was models were not able to find a opposing trend between the directions.
Vii-B2 Power Spectral Density and Fourier Transform
Finally we get to the Power Spectral Density/Frequency Domain analysis to figure out what exactly was the models classifying. Power Spectral Density (PSD) is a frequency-domain plot of power/Hz vs frequency. Power Density tells us which frequencies contain the signal’s power, measure in amplitude2̂/Hz. We employed Fourier Transform to decompose the data into sinusoidal components. Then, we used the components to further calculate the amplitude at each frequency.
If we look at the frequency domain visualization of the different direction of wind in Fig 5. We see that for no wind, front and back wind the amplitude of the power spectrum is around 400 to 500 for roll data. While for left and right wind the amplitude is much higher. This explains the miss classifications, as if we ignore the higher amplitude of data, were only small sets of data points lie, we notice that there is a large number of data points that lie within the same amount of change from 0 power/Hz (the equilibrium point) to 500 power/Hz for all directions. Thus we can conclude for this study, the classification models were classifying the amplitude of the sinusoidal components in data and because of this, with many data point for all directions lying in the same deviation from equilibrium point, the classifier will miss classify the data points.
Viii Supplemental Analysis and Testing
For study 1 we also applied Fast Fourier Transformation VIII-A
to our data to test if it will increase our accuracy and in study 2, we used a Convolutional Neural NetworkVIII-Bto test our raw dataset.
Viii-a Fast Fourier Transform (FFT)
FFT acts as a filtering mechanism after analyzing the frequency domain of the dataset. We determine our cutoff point for FFT in the frequency domain by taking the k most dominant number of data points from the 100 data point time window, XV. The Table XVI shows the results of applying FFT to our raw dataset of Drone 1, with marginal increase in accuracy, especially at the cost of increase in computational intensity. Furthermore, when the dataset became large enough, there was no increase in accuracy observed when applying FFT. However, for smaller dataset like when dealing with real-time analysis, when we are only able to analyze 100 data point at once, FFT could prove valuable.
Drone1, Random Forest Classifier
|No FFT Applied||FFT Applied, k = 30|
Viii-B Neural Network, Convolutional Neural Network (CNN) with Tensorflow
In Study 2, we explored the use of Convolutional Neural Network (CNN) as it allows the model itself to choose the most optimal features from raw data. The CNN model consists of one convolution layer followed by max pooling and another convolution layer. After that, the model will have fully connected layer which is connected to Softmax layer. The model was based our model off of human activity recognition cnn models222http://aqibsaeed.github.io/2016-11-04-human-activity-recognition-cnn/. However the results from our initial analysis with CNN left much to be desired, which can be seen in the Fig 6 showing an accuracy of just 24.87%.
In this paper, we propose a new methodology for wind detection in quadcopters (drones). Our proposed approach has demonstrated its instrumentality in within-drone and across-drone wind speed detection, making it possible to apply such detection in real-time. We believe this framework will be useful for MAVs/UAVs when operating in areas with high air turbulence, helping them avoid fatal crashes due to strong wind gusts, without the need to add additional bulky sensors. This will also help reduce weight thus increasing battery life of drone flight which is very much limited at this stage of the technology. This project has set a great benchmark for further work in the future in this area of testing. Especially if we can alter the features for direction detection by futher research into Deep Neural Networks and better raw dataset. It is to our knownledge that the Fordham Robotics Lab recently implemented ROS(Robot Operating System) drivers for Crazyflie 2.0 drones, which we had no access to before. This will allow us to collect more specific data and manipulate the drone in a more advanced ways.
X Thank You
Finally we would like to thank Dr. Li Yanjun for all the help and guidance during this capstone project course, raising valuable questions about our methodologies, so as to help us find the right answers to our problems. We would also like to thank Dr. Lyons, Dr. Nguyen and the Fordham Robotics Lab for their support, equipment and guidance throughout this project.
N. Ravi, N. Dandekar, P. Mysore, and M. L. Littman, “Activity recognition from
accelerometer data,” in
Proceedings of the 17th conference on Innovative applications of artificial intelligence-Volume 3. AAAI Press, 2005, pp. 1541–1546.
-  G. M. Weiss, J. L. Timko, C. M. Gallagher, K. Yoneda, and A. J. Schreiber, “Smartwatch-based activity recognition: A machine learning approach,” in Biomedical and Health Informatics (BHI), 2016 IEEE-EMBS International Conference on. IEEE, 2016, pp. 426–429.
-  J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Cell phone-based biometric identification,” in Biometrics: Theory Applications and Systems (BTAS), 2010 Fourth IEEE International Conference on. IEEE, 2010, pp. 1–7.
-  A. H. Johnston and G. M. Weiss, “Smartwatch-based biometric gait recognition,” in Biometrics Theory, Applications and Systems (BTAS), 2015 IEEE 7th International Conference on. IEEE, 2015, pp. 1–6.
-  F. Guo, Y. Li, M. S. Kankanhalli, and M. S. Brown, “An evaluation of wearable activity monitoring devices,” in Proceedings of the 1st ACM international workshop on Personal data meets distributed multimedia. ACM, 2013, pp. 31–34.
-  J. Engel, J. Sturm, and D. Cremers, “Accurate figure flying with a quadrocopter using onboard visual and inertial sensing,” Imu, vol. 320, p. 240, 2012.
-  S. Ross, N. Melik-Barkhudarov, K. S. Shankar, A. Wendel, D. Dey, J. A. Bagnell, and M. Hebert, “Learning monocular reactive uav control in cluttered natural environments,” in Robotics and Automation (ICRA), 2013 IEEE International Conference on. IEEE, 2013, pp. 1765–1772.
-  T. Mori and S. Scherer, “First results in detecting and avoiding frontal obstacles from a monocular camera for micro unmanned aerial vehicles,” in Robotics and automation (icra), 2013 ieee international conference on. IEEE, 2013, pp. 1750–1757.
-  W. G. Aguilar, V. P. Casaliglla, and J. L. Pólit, “Obstacle avoidance based-visual navigation for micro aerial vehicles,” Electronics, vol. 6, no. 1, p. 10, 2017.
A. Benini, A. Mancini, and S. Longhi, “An imu/uwb/vision-based extended kalman filter for mini-uav localization in indoor environment using 802.15. 4a wireless sensor network,”Journal of Intelligent & Robotic Systems, vol. 70, no. 1-4, pp. 461–476, 2013.
K. McGuire, G. de Croon, C. De Wagter, K. Tuyls, and H. Kappen, “Efficient optical flow and stereo vision for velocity estimation and obstacle avoidance on an autonomous pocket drone,”IEEE Robotics and Automation Letters, vol. 2, no. 2, pp. 1070–1076, 2017.
-  M. Rossi, D. Brunelli, A. Adami, L. Lorenzelli, F. Menna, and F. Remondino, “Gas-drone: Portable gas sensing system on uavs for gas leakage localization,” in SENSORS, 2014 IEEE. IEEE, 2014, pp. 1431–1434.
-  M. Rossi and D. Brunelli, “Autonomous gas detection and mapping with unmanned aerial vehicles,” IEEE Transactions on Instrumentation and Measurement, vol. 65, no. 4, pp. 765–775, 2016.
-  N. Sydney, B. Smyth, and D. A. Paley, “Dynamic control of autonomous quadrotor flight in an estimated wind field,” in Decision and Control (CDC), 2013 IEEE 52nd Annual Conference on. IEEE, 2013, pp. 3609–3616.
-  E. J. Smeur, G. C. de Croon, and Q. Chu, “Gust disturbance alleviation with incremental nonlinear dynamic inversion,” in Intelligent Robots and Systems (IROS), 2016 IEEE/RSJ International Conference on. IEEE, 2016, pp. 5626–5631.
-  J. Paneque-Gálvez, M. K. McCall, B. M. Napoletano, S. A. Wich, and L. P. Koh, “Small drones for community-based forest monitoring: An assessment of their feasibility and potential in tropical areas,” Forests, vol. 5, no. 6, pp. 1481–1507, 2014.
-  V. V. Klemas, “Coastal and environmental remote sensing from unmanned aerial vehicles: An overview,” Journal of Coastal Research, vol. 31, no. 5, pp. 1260–1267, 2015.
-  L. Tang and G. Shao, “Drone remote sensing for forestry research and practices,” Journal of Forestry Research, vol. 26, no. 4, pp. 791–797, 2015.
-  C. G. Knight, “Low airspeed measuring devices for helicopter usage monitoring,” DEFENCE SCIENCE AND TECHNOLOGY ORGANISATION VICTORIA (AUSTRALIA) PLATFORM …, Tech. Rep., 2003.
-  Bitcraze AB, “Crazyflie 2.0,” 2016. [Online]. Available: https://www.bitcraze.io/crazyflie-2/