Publications

@mastersthesis{uuid:82cb0f68-061e-4346-b536-a35a61621e51,

title = {Detecting Empty Wireframe Objects on Micro-Air Vehicles: Applied for Gate Detection in Autonomous Drone Racing},

author = {Philipp Dürnay},

url = {http://resolver.tudelft.nl/uuid:82cb0f68-061e-4346-b536-a35a61621e51},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Electrical Engineering, Mathematics and Computer Science},

abstract = {Autonomous MAV are an emerging technology that supports a wide range of applications such as medical delivery or finding survivors in disaster scenarios. As flying in such missions is difficult the robust estimation of an MAV's state within its environment is crucial to ensure safe operation. In indoor scenarios, cameras are one of the predominant choices for state estimation sensors. This requires Computer Vision algorithms to interpret the obtained high dimensional signal. An application that allows the competitive evaluation of control and state estimation algorithms is MAV Racing such as the IROS 2018 Autonomous Drone Race. Thereby a race court consisting of several race gates has to be followed. For a fast flight during such a race court the detection of the racing gates with a camera can be used in a high level control loop. As these objects consist only of small structures that are spread across large parts of the image, this gives rise to a challenging Object Detection problem. In recent years CNN showed promising results on various vision tasks. However, due to their computational complexity the deployment on mobile devices remains a challenge. Furthermore, CNN typically require a vast amount of training data. Finally, the objects typically studied in Object Detection consist of solid and complex features which is not the case for racing gates. Therefore, this work defines the class of EWFO and studies their detection on MAV with YoloV3. Thereby, the training data is created with a graphical engine. We are interested in how to detect EWFO with a CNN on a MAV, using synthetic data. We conduct several simple experiments about EWFO in simulation and compare their detection to more filled objects. Subsequently experiments in a more challenging environment such as an MAV race are conducted. The experiments show how EWFO are harder to detect than filled objects as the detector can be confused to patterns present in the empty part. Particularly for larger objects the detection performance decreases. We give several recommendations on how to generate data for the detection of EWFO on MAV. These include how to add variations in background as well as the camera placement. Finally, we study the incorporation of image augmentation techniques to transfer the detector to the real world. We can report that especially modelling lens distortion improves the performance on the real data. Nevertheless, a reality gap remains that can not fully be explained. Furthermore, different architectures are studied for the detection of EWFO. It can be seen how a relatively shallow network of 9 layers can be used for the detection of EWFO on MAV. A further reduction in weights leads to a gradual decrease in performance. Based on the gained insights the deployment of a detector on the example system JeVois is studied. A detection performance/speed trade-off is evaluated. The final detector achieves 32% average precision at a frame rate of 12 Hz on a real world test set created during this work. The gained insights can be used to deploy the detector in a control loop for MAV. This ensures the safe flight through a racing court of an autonmous drone race. The gained insights about the detection of EWFO can be transferred to objects with similar properties},

note = {Tax, D.M.J. (mentor); de Croon, G.C.H.E. (mentor); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{uuid:bc31203c-1956-4665-a92a-8203881f22ce,

title = {Attitude modeling of the DelftaCopter: a system identification approach},

author = {Joost Meulenbeld},

url = {http://resolver.tudelft.nl/uuid:bc31203c-1956-4665-a92a-8203881f22ce},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Aerospace Engineering; TU Delft Control & Simulation},

abstract = {Previous years have seen a rise in the use of Unmanned Aerial Vehicles (UAVs). Reaching a large endurance and range while being able to perform Vertical Take-Off and Landing (VTOL) landings allows a broad range of applications. For this purpose the DelftaCopter (DC) was developed, a tilt-body tailsitter UAV. It hovers using a single helicopter rotor for lift and transitions to forward flight by pitching its body down by 90°. In this forward flight state, wings generate the lift, while the helicopter rotor now provides thrust. The single rotor is more efficient than using multiple smaller rotors and helicopter swashplate is used for attitude and speed control. The heavy single helicopter rotor introduces significant gyroscopic moments, as is the case for all helicopters. In contrast with normal helicopters, the DC has a heavy fuselage putting the attitude dynamics between a helicopter and aircraft. In previous research, a controller based on a model incorporating the rotor as a rotating cylinder was implemented. This controller was unable to counteract the gyroscopic pitch-roll coupling, leading to the question of this thesis: how should the DC be modeled to allow control design. <br/>In this thesis, the previous model is called the Cylinder Dynamics (CD) model, and is compared with another model from literature. The latter model, in this thesis called the Tip-Path Plane (TPP) model, includes the flapping dynamics through the tip-path plane dynamics and is also a linear state-space model. In flight tests, chirps were used to cover a broad frequency range. Fitting both the CD and TPP models on this flight test data, it is shown that the CD model lacks accuracy in the high-frequency area, while the TPP is able to accurately model these dynamics. This shows that the flapping dynamics are important to the attitude dynamics of the DC. An Linear Quadratic Regulator (LQR) controller was implemented based on the fitted TPP model, and shows adequate tracking performance, further validating the applicability of the model to the DC. For forward flight, extensions to the hover models are proposed. The extension including the elevator and aerodynamic damping is shown to simulate key dynamics of the DC in forward flight with reasonable accuracy. The parameters and eigenfrequencies of this model are not significantly different from the hover model. Therefore it can be concluded that the gyroscopic effect plays an important role in forward flight attitude dynamics. Another extension which estimates of angle of attack and sideslip using high-pass filtered rotational rates, yields better accuracy, but significantly changes the model parameters also present in the hover model. More research with angle of attack and sideslip vanes could validate this modeling approach. It was also found that for a new version of the DC with a smaller, more quickly rotating rotor, the modeling done before resulted in much worse fits. It was shown that the CD and TPP model response is much more comparable for this version. Control performance also suffers due to this lower accuracy model fit. Further research is required to understand why this is the case.},

note = {de Wagter, C. (mentor); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{uuid:28fad2a0-4b4c-47f4-8930-01708f4b52d1,

title = {Hear-and-avoid for UAVs using convolutional neural networks},

author = {Dirk Wijnker},

url = {http://resolver.tudelft.nl/uuid:28fad2a0-4b4c-47f4-8930-01708f4b52d1},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Aerospace Engineering},

abstract = {We investigate how an Unmanned Air Vehicle (UAV) can detect manned aircraft with a single microphone. In particular, we create an audio data set in which UAV ego-sound and recorded aircraft sound can be mixed together, and apply convolutional neural networks to the task of air traffic detection. Due to restrictions on flying UAVs close to aircraft, the data set has to be artificially produced, so the UAV sound is captured separately from the aircraft sound. The aircraft data set is collected at Lelystad airport by capturing flyovers with a microphone array. It is mixed with UAV recordings, during which labels are given indicating whether the mixed recording contains aircraft audio or not. The mixed recordings are the input for a model that determines whether an aircraft is present or not. The model is a CNN which uses the features MFCC, spectrogram or Mel spectrogram as input. For each feature the effect of UAV/aircraft amplitude ratio, the type of labeling, the window length and the addition of third party aircraft sound database recordings is explored. The results show that the best performance is achieved using the Mel spectrogram feature. The performance increases when the UAV/aircraft amplitude ratio is decreased, when the time window is increased or when the data set is extended with aircraft audio recordings from a third party sound database. It is not desirable to train the model on distant approaches and test them on nearby approaches as the performance then drops. The results also prove that the performance increases the closer the aircraft is. Although the currently presented approach has a number of false positives and false negatives, that is still too high for real-world application, this study indicates multiple paths forward that can lead to an interesting performance. In addition, the data set is provided as open access, allowing the community to contribute to the improvement of the detection task.},

note = {van Dijk, Tom (mentor); de Croon, G.C.H.E. (mentor); de Wagter, C. (mentor); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{uuid:b0394f21-302c-484d-a6dc-031f5860c521,

title = {Flight control and collision avoidance for quadcopter and flapping wing MAVs using only optical flow: Theory, Simulation and Experiment},

author = {Daan Vrede},

url = {http://resolver.tudelft.nl/uuid:b0394f21-302c-484d-a6dc-031f5860c521},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Mechanical, Maritime and Materials Engineering},

abstract = {Both quadcopter Micro Aerial Vehicles (MAVs) and Flapping Wing MAVs (FWMAVs) are constrained in Size, Weight and Processing power (SWaP) in order to achieve flight tasks that include attitude and velocity stabilisation, as well as obstacle avoidance. <br/>Conventional sensory and control approaches, such as those relying on inertial, visual and Global Positioning System (GPS) sensors, can fulfil these tasks using sensor fusion. However such approaches do not score well in terms of SWaP criteria. <br/>Very simple proportional feedback control laws using single optical flow vectors from very basic high frame-rate low-resolution cameras provide a promising path to achieve aforementioned tasks. <br/>This thesis shows that in theory these control laws are well suited for stabilising a FWMAV, and could be used for a high-drag adapted quadcopter MAV within bounds. Simulations confirm these findings and illustrate robustness to noise and additional emergent behaviour such as sideways wall avoidance and trajectory following, however simulations also show that disparity between walls can lead to unintended rotational behaviour during vertical translation. <br/>The system is tested in experiment on a quadcopter-like setup with onboard processing, using only ADNS 9800 computer mouse optical flow sensors for flight control. Results show that the system behaves similarly to simulation, however the sensory configuration used is highly dependent on texture in environment and light conditions. <br/>For future work it is recommended to investigate optical flow sensors in more detail to obtain reliable output on a vibrating platform (such as a FWMAV) in a broader range of texture and light conditions. Preliminary results from theory, simulation and experiment indicate that the addition of derivative feedback could strongly enhance performance on a quadcopter MAV and remove the requirement for high drag.},

note = {Goosen, J.F.L. (mentor); de Croon, G.C.H.E. (mentor); Breedveld, P. (mentor); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{uuid:c75ab7d9-e711-4e0c-93ea-ff092e2e9131,

title = {Estimation of ego-motion velocities from single static images},

author = {Yeshwanth Napolean},

url = {http://resolver.tudelft.nl/uuid:c75ab7d9-e711-4e0c-93ea-ff092e2e9131},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Aerospace Engineering},

abstract = {Velocity estimation based on visual information is a well- researched topic. Traditional approaches usually rely on how a given feature or features change between successive images in a sequence. However, a single static image might contain motion information that could potentially be lever- aged to estimate the optical flow. It can be hypothesized that motion blur and context of the scene are two sources of mo- tion information in static images. This research work has two main goals, one is to investigate the prospect of using a learning-based framework to model a mapping directly to camera ego-motion velocity. The second goal is to ana- lyze the contributing features in learning such a mapping. Experiments show that the model is able to learn velocity based on context of the scene but performs better when in- put images contain motion blur.<br},

note = {de Croon, G.C.H.E. (mentor); van Gemert, J.C. (graduation committee); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{uuid:82cfe98f-46ad-42b4-8021-082fbae1f740,

title = {Passive Localization of Robots with Ambient Light},

author = {Danielle Werff},

url = {http://resolver.tudelft.nl/uuid:82cfe98f-46ad-42b4-8021-082fbae1f740},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Electrical Engineering, Mathematics and Computer Science},

abstract = {A lot of research is been being done on Visible Light Communication (VLC), which has shown to be of interest for many applications, such as localization. Since localization based on VLC requires active modulation of light sources, this limits the amount of light sources that can be used for localization. Furthermore, in some situations there might not even be a controllable light source present (for example outdoors). To extend the use of light-based localization schemes, this thesis looks into a way to achieve the same result as current VLC localization methods in a passive manner, i.e. without control of the light sources. <br/><br/>Previous work has been done on passive ambient light-based localization by Wang et al.: objects are equipped with unique barcodes, that reflect ambient light in a distinct manner. The reflected light is received by photosensors, from which their ID is obtained. However, this work has focused on identifying large-sized objects in one dimension. Using the same principle for localization of small-sized objects, and in two dimensions, are open challenges that this thesis addresses . <br/><br/>The work presented here forms a proof-of-concept of a passive light-based localization system for two-dimensional, real-time tracking of small-sized objects. In order to achieve this, a special enclosure has been designed, giving simple photosensors the ability to distinguish small-sized objects without compromising their FOV. With this enclosure, a single photosensor can detect barcodes down to 7 cm in size in the test set-up, while distinguishing up to three different IDs. A particle filter has been implemented to combine detections from different photosensors into a single estimate of an object’s location. <br/><br/>The localization system is designed around the robots designed by a MSc student at the Embedded Systems group at TU Delft. By moving these robots at a speed of 15.4 cm/s in a straight line through the test set-up, a localization error of 4.8 cm is obtained. The distance between the robots and the sensor equals 20 cm.},

note = {Zuñiga Zamalloa, Marco (mentor); Pawelczak, Przemek (mentor); de Croon, Guido (graduation committee); Langendoen, Koen (graduation committee); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{uuid:d3bd3ba9-61aa-4d48-b13d-aead118b8015,

title = {Vector Field Based Path Following for UAVs using Incremental Nonlinear Dynamic Inversion},

author = {Suresh Sharma},

url = {http://resolver.tudelft.nl/uuid:d3bd3ba9-61aa-4d48-b13d-aead118b8015},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Aerospace Engineering; TU Delft Control & Operations},

abstract = {This work presents a vector field based path following method to be used by Multirotor Unmanned Aerial Vehicles (UAVs). The desired path to be followed is a smooth planar path defined in its implicit form. The vector field around the desired path is then constructed using the implicit function, such that the integral curves of the vector field converge to the path. The algorithm takes into account the future change in the trajectory as well as the current state of the UAV in order to calculate the desired linear acceleration, which is then tracked using the Incremental Nonlinear Dynamic Inversion (INDI) controller in the autopilot. The implementation also allows for the velocity of the UAV to be controlled independently. The efficiency of the algorithm is demonstrated using real world flight tests, and the performance is shown to be better than the<br/>traditional carrot-chasing controller.},

note = {Smeur, E.J.J. (mentor); Chu, Q. P. (mentor); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{uuid:3794f912-f141-4fa7-aff4-464598958e94,

title = {Vision-based Autonomous Drone racing in GPS-denied Environments},

author = {Michaël Ozo},

url = {http://resolver.tudelft.nl/uuid:3794f912-f141-4fa7-aff4-464598958e94},

year  = {2018},

date = {2018-01-01},

school = {TU Delft Aerospace Engineering},

abstract = {High-speed autonomous flight of Micro Air Vehicles has gained much attention in recent years. However, flight in complex GPS-denied environments still poses a serious challenge. One scenario which contains these elements is drone racing, where pilots have to fly complex tracks at high speed, often in an indoor environment. In this work we therefore present an MAV capable of autonomously flying such a drone race track. The system has to operate in a GPS-denied environment, hence a visual navigation method is employed. Position is recovered from gate detections based on a novel least-squares method, while heading is estimated using an optimization based method. Experiments show that both methods have a higher accuracy than the standard P3P pose estimation method. Furthermore, a state estimation filter is designed to fuse the visual measurements with IMU measurements, by using an EKF with drag based prediction model. For high-level control different motion primitives are linked, which allow the MAV to fly the track without having a detailed on-board map. The overall approach does not rely on SLAM or Visual odometry, which results in low computational complexity. Also, it does not rely on downward optical flow velocity measurements, which enables it to work even in low texture environments.},

note = {de Croon, G.C.H.E. (mentor); Delft University of Technology (degree granting institution)},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}