Examensarbeten för masterexamen // Master Theses
Länka till denna samling:
Browse
Browsar Examensarbeten för masterexamen // Master Theses efter Program "Complex adaptive systems (MPCAS), MSc"
Visar 1 - 20 av 62
Sökresultat per sida
Sortera efter
- Post3-D object tracking through the use of a single camera and the motion of a driverless car(2021) Ovnell, Andreas; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Benderius, Ola; Benderius, OlaThere has been a very large increase in interest and development of partially or fully driverless cars in recent years. For these driverless cars to function, they need to be able to navigate to their destination while avoiding nearby objects. This can be done using simultaneous localisation and mapping (SLAM). SLAM is the task of simultaneously creating a map of the surrounding objects while keeping track of the car’s position within this map. This thesis will look into the feasibility of using a single camera attached on a driverless car to perform SLAM on cones detected by the real-time object detection system You only look once (YOLO). Three different methods were tested. All of these require a calibrated camera that is capable of determining horizontal and vertical angles from the pixel positions. The first ‘triangulation’ method uses that the distance travelled and rotation between two frames is known. The second ‘plane projection’ method is an optimisation problem which consists of finding the variables which result in lowest error, and through this determine the cone distances and car speed. The map of the surrounding cones is moved according to the estimated velocity and rotation of the car such that the car is always placed at the origin, allowing for use of multiple detections to improve accuracy. The third ‘distance from cone height’ method works by using the size of the cone detections in order to determine the approximate distance of each cone, use this to determine the approximate angle of the camera and then use the median angle to make the final distance estimates. The triangulation method was shown to be completely unsuitable for mono-camera use. The plane projection method was shown to be unreliable, likely due to a relatively small number of visible cones and a too large noise amplitude of detections from YOLO. The distance from cone height method was shown to be the best out of the tested methods, as it was simple, fast and quite reliable. However, this method still had an error approximately 1.4 times larger than what is advertised by commercial stereo camera systems.
- PostA new generation humanoid robot platform(2011) Magnus, Wahlstrand; Chalmers tekniska högskola / Institutionen för tillämpad mekanik; Chalmers University of Technology / Department of Applied MechanicsThere are many tasks that humans for different reasons are unwilling or unfit to do. Examples are, for instance, dangerous tasks such as handling toxic waste or monotonous tasks, like working in assembly lines. The hope is that robots one day can do tasks like these for us. Even though a lot of progress has been made in robotics in the last few decades, it is clear that a lot of work and research remains for this goal to be fulfilled. This master thesis describes a process of upgrading Kondo, a small humanoid robot, from a basic robot with no sensory capabilities to a more advanced robotic platform. The hope is that the improved platform can be used to facilitate further research in several fields of robotics such as human-robot interaction, adaptive control and evolutionary robotics. In order to perform this upgrade, the servo controller of existing platform was replaced by a new programmable servo controller. Furthermore, a sensor module with an accelerometer and distance sensors was designed and added to the platform, giving Kondo sensory capabilities. To complete the system, a two part software interface was created. This included a graphical user interface to directly control the robot and create motion sequences and a Python class interface for prototyping and more advanced programs. The resulting platform was tested in order to ensure that it fulfilled the objectives stipulated in the project. The tests included hardware testing, i.e. testing the actual motion of the robot and the communication between to and from the electronic modules. The platform’s configurability was also tested by implementing three common robotic features, including automated fall recovery and wall avoidance. The results of these tests indicate that the basic functionality of the new platform, such as walking and standing, is rather robust. The speed of the developed gait however, can be improved. The platform is relatively easy to extend and modify therefore can be used in education or in robotic research. A weakness of the current platform is the number of connections needed to power and communicate with the electronic boards. Decreasing this number is something that could be worked on in future projects in order to increase the robot’s autonomy.
- PostAnomaly Detection in Logged Sensor Data(2015) Florbäck, Johan; Chalmers tekniska högskola / Institutionen för tillämpad mekanik; Chalmers University of Technology / Department of Applied MechanicsAnomaly detection methods are used in a wide variety of fields to extract important information (e.g. credit card fraud, presence of tumours or sensor malfunctions). Current anomaly detection methods are data- or application specific; a general anomaly detection method would be a useful tool in many situations. In this thesis a general method based on statistics is developed and evaluated. The method includes well-known statistical tools as well as a novel algorithm (sensor profiling) which is introduced in this thesis. The general method makes use of correlations found in complex sensor systems, which consists of several sensor signals. The method is evaluated using real sensor data provided by Volvo Car Corporation. The sensor profiling can be used to find clusters of data with similar probability distributions. It is used to automatically determine the sensor performance across different external conditions. Evaluating the anomaly detection method on a data set with known anomalies in one sensor signal results in 94 % of anomalies detected at 6 % false detection rate. Evaluating the method on additional sensor signals was not done. The sensor profiling revealed conditions where the sensor signal behaves qualitatively and quantitatively different. It is able to do this in data where other commonly used methods, such as regression analysis, fail to extract any information. Sensor profiling may have additional applications beyond anomaly detection as it is able to extract information when other methods can not. To conclude, this thesis presents a seemingly natural method and tool chain to automatically detect anomalies in any sensor data that can be represented as a time series. The performance of this method is still to be proven on a large set of general sensor data, but it shows promise, mainly for sensor systems consisting of several sensor signals.
- PostArtificial intelligence based marine autopilot: Trained using reinforcement learning in the Unity simulation environment(2022) Asplund, Martin; Näslund, David; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Forsberg, Peter; Ahlstedt, MikaelWhen driving a ship, especially long distances, an autopilot is usually deployed. Since the ship is subjected to disturbances from currents and wind alike, it is not easy to keep a steady course. Furthermore, the dynamic behaviour of waves striking the vessel, known as swaying, makes the task of keeping a straight line through the sea daunting. To assist, there exist commercial autopilots. However, most of these are subject to less than simple calibration, which also is hard to keep accu rate throughout the lifespan of the boat due to wear, different load conditions, and other similar things. Also, there is generally no adaptability related to the autopilot, meaning a sudden change in engine performance will stop the autopilot from functioning. Further, the majority of today’s commercial autopilots are designed to follow a course or a heading, known as course-hold and heading-hold autopilots. Hence, there exists a desire to develop a more adaptable path-following autopilot. One way of solving the adaptability issue is to borrow the solution from the aircraft industry and use a control allocator. Given a set of global forces (usually Fx, Fy, and moment Mz) the control allocator tries to distribute these between a given set of actuators. Since the number of control signals usually is far less than the number of actuators, these systems are said to be over-actuated, and no unique solution exists. This work aims at exploring a new way of constructing computationally efficient regulation and control allocation for vessels, in the form of a path-following autopilot. The hypothesis is that, by using a Neural Network as a control allocator, better performance and adaptability than offered by present solutions can be achieved.
- PostAssessment and development of robust software for communication in buses(2015) Stenberg, Jim; Chalmers tekniska högskola / Institutionen för tillämpad mekanik; Chalmers University of Technology / Department of Applied MechanicsNew quality demands combined with existing problems with software updates led to a desire of improved robustness in one of Volvo Buses’ control units. The code was analyzed by a static code analysis tool, tested and manual reviewed. A breakdown into five different fields (writing convention for the code, incorrect input, internal error, memory shortage and work overload) allowed for focused analysis. The gathered results were then used to design improvements for the unit, some implemented and some only evaluated. The aim was to increase the unit’s robustness when communicating. The changes made improved the memory usage and calculated the work load more accurately. This led to less resets and a more robust system. However, if more robustness would be needed in the future, the current operating system would need to be replaced.
- PostAutomated Detection of Pain in Horses through Facial Expression Analysis(2016) Bhatti, Kashif; Chalmers tekniska högskola / Institutionen för tillämpad mekanik; Chalmers University of Technology / Department of Applied MechanicsA method for automated pain-assessment in horses through facial-expression analysis is proposed. The method is based on supervised linear classification of a feature stack of Gabor filters and has the desirable quality of not requiring expert knowledge or specialized equipment to make an assessment. The method is evaluated by applying it to images of horses from two clinical trials where the horses were (ethically) subjected to pain. The resulting accuracy of 78% compares favorably to an alternate method of pain assessment based on facial expression cues that requires expertise to administer.
- PostAutomated semantic grammar generation in dialogue-based task-oriented systems(2021) Triantafyllou, Georgios; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Wahde, Mattias; Wahde, MattiasDialogue-base task-oriented systems are conversational systems that can complete tasks and answer questions. A semantic grammar can be helpful for user input recognition in such systems. Manually creating semantic grammars can be cumbersome for a developer and delay the deployment of the system, thus it would be of great assistance to automate the process. The topic of automated semantic grammar generation has not been deeply explored and current approaches in this field are either semi-automated or lack interpretability and robustness. In this thesis, the prospect of developing a completely automated semantic grammar generation method was investigated. An additional aim was to create a method that could be employed in different domains without needing any additional modifications. We propose a novel method that utilizes syntactic and relation analysis to infer the semantics of a user input. The method developed is a hybrid approach that comprises statistical methods for the syntactic analysis and a rule-based model for the semantics. The results show that the method is able to generate a fairly accurate and precise semantic grammar. It is also observed that the method is not able to fully analyze all possible pattern cases. Furthermore, the relation analysis has proved to be helpful on finding semantic synonymity amongst different user input patterns. It is plausible to completely automate the creation of a semantic grammar and our findings suggest that hybrid approaches can preform well in this task. Nevertheless, further adjustments should be made in the rule-based model to provide a universal coverage of pattern cases
- PostAutonomous vehicle control: Exploring driver modelling conventions by implementation of neuroevolutionary and knowledge-based algorithms(2018) Arnö, Linus; Eriksson, Jonas; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Chalmers University of Technology / Department of Mechanics and Maritime SciencesIn this paper an investigation of driver modelling conventions is presented. The goal was to compare traditional driver modelling with machine learning, to nd indications of when one approach could be preferred over the other. This was done by implementing some representatives of the di erent approaches and evaluating them in the same conditions. The traditional approach was represented with one well established model by Sharp et al., as well as one self made aim point model. Both of these required a path planner and velocity control, that were also designed by the authors themselves. The machine learning approach was represented by neuroevolution, an alternative technique for solving reinforcement learning problems, and speci cally the method called NEAT. The results showed that all implemented methods were able to solve the task, but in the speci c scenario and with the current amount of training the two traditional models were superior to the evolved neural network. Similarities and potential reasons for di erences between the models are discussed, as well as some identi ed advantages and disadvantages to both approaches.
- PostBrain–computer interface for autonomous vehicles: An investigation of a hands-free control system(2019) Kowalska, Kamila; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Benderius, Ola; Berger, ChristianBrain–computer interfaces have been a subject of growing interest in recent years and devices measuring brainwave activity are being used within vehicles for evaluating attentiveness and detecting steering intentions. Implementation of such an interface for vehicle control is the main focus of this thesis. In this work, a novel EEG-based interface supported by eye tracking is investigated in a vehicle control application. The proposed solution merges algorithms utilized in autonomous driving systems with a brain–computer interface based on a P300 response. The control method introduces target following rather than directional steering as a principle of BCI driving, potentially simplifying control and reducing the influence of delay typical for electroencephalography classification. The interface has been tested by five untrained participants in a simulated laboratory environment. The testing platform consisted of an OpenBCI EEG headset, Pupil Labs wearable eye tracker connected to a standard PC unit, and a miniature robot platform equipped to semi-autonomously maneuver, follow and avoid objects which served as a controlled vehicle. The participants were asked to perform simple driving tasks by observing the frontal camera feed on the computer monitor while their brain response was being recorded and a signal pattern acted as a trigger. Target marking was realized by tracking the gaze position in the character of a selector, and a brain response matching a theoretical P300 was interpreted as a will to interact with the object. The subjects were interacting with the interface intuitively and were generally able to complete the tasks. The hardships arose in relation to the measuring equipment, which was revealed to be of unsatisfactory quality. While the results are very promising and point to the proposed target-based steering as preferable for BCI driving, further work is recommended to fully estimate the applicability in a real world scenario.
- PostBuilding a Computer Vision System for Autonomous Tree Planting: Using YOLO and U-Net to Find Planting Spots in Clear-Felled Areas(2022) Christenson, Olle; Lundgren, Jens; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Wolff, Krister; Wolff, KristerPlanting new saplings at clear-felled areas is an expensive and demanding task for forest owners. To improve efficiency and consistency, Södra Skogsägarna has initiated a project to develop an autonomous vehicle to plant new saplings in clear-felled areas. A crucial function of the system is how to select the planting spots. This thesis aims to create a deep learning-based computer vision model to locate favorable planting spots. A stereo camera that provides RGB-D data from different scenes, where a sapling should be planted, will be used. The created model takes this data as input and returns the coordinates of two proposed planting spots. The model is based on a YOLO network for object detection and two different implementations of U-Net networks for segmentation. The algorithm was able to find good planting spots in 81% of the test cases. A discussion of the most common reasons why the model occasionally proposes invalid planting spots and suggestions on how to solve these problems are given. Suggestions are also given on how the project group could proceed with the project and improve the system. The main conclusions are that a better suited camera than the one used in this thesis should be used and that more data should be collected to increase the models robustness. The code for the final system can be found in the repository https://github.com/Birken666/Master-Thesis.
- PostCollective transportation of objects by a swarm of robots(2015) Torabi, Sina; Chalmers tekniska högskola / Institutionen för tillämpad mekanik; Chalmers University of Technology / Department of Applied MechanicsA collective transport strategy, inspired by the food retrieval procedure of ant colonies, has been implemented on a swarm of robots that are smaller than the object. A simple odometry-based team coordination strategy in combination with an omni-directional camera has been implemented, resulting in a well-coordinated e ort by the robots without using any communication. The strategy is fully decentralized. Moreover, a simple recruitment process has been introduced but it did not improve the transportation efficiency. The transportation strategy consists of four stages, namely prey discovery, team coordination, recruitment, and transportation. A simulation environment capable of handling robot swarms and their physical interaction is developed for this project. Using robots weighing 3 kg, a 3 kg object was successfully transported in 48 out of 50 trials, whereas a 4.5 kg object was successfully transportedin 44 out of 50 trials.
- PostConcept exploration of spacecraft separation structures using evolutionary optimization(2022) Arnebro Söderberg, Erik; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Wahde, Mattias; Wahde, MattiasConventional concept exploration tends to be narrowly focused around previous solutions, especially in an industry with as much emphasis on reliability as aerospace engineering. A drawback of this approach is the likelihood of high performing design concepts existing outside that narrow design domain. In this thesis project a tool is developed which aims to augment the concept exploration phase by autonomously searching a broad range of feasible solutions. The search converges to the non-dominated set by means of multi-objective evolutionary optimization, wherein the structures are optimized for lower mass and higher natural frequency. The candidate solutions are modelled, meshed, and structurally analysed using open-source software. The tool was tested on two different types of structures, namely launch vehicle adapters (LVA) and satellite dispenser structures. LVA design solutions with both higher natural frequency (up to 4% increase) and reduced mass were discovered. The results show that the developed method is able to converge to a set of high performing solutions, and present these to the user as input for subsequent design decisions.
- PostCuriosity based Self-Organization of Humanoid Robot(2015) Loviken, Pontus; Chalmers tekniska högskola / Institutionen för tillämpad mekanik; Chalmers University of Technology / Department of Applied MechanicsThis thesis presents a novel approach to how a high dimensional humanoid robot of 18 dimensions can learn within a few hours to control its body so that it is able to perform simple tasks such as rolling around or to sit up. The method is robust and works equally well when an arm is removed, and in a case where the robot was trained to use two arms and one was removed it quickly adapted to its new body. The robot is equipped with an accelerometer that measures the tilt of the torso in 2 dimensions. This "tilt"-space is divided into a discrete set of states, and the way in which the dimensionality of the servo-space is made irrelevant is to only allow one servo-con guration per state. These con gurations are evolved using a Self-Organizing Map, while an Arti cial Curiosity-driven Reinforcement Learner chooses what state to state transitions to attempt. An additional parameter is added in a nal experiment, to see if the agent can even learn to stand. This experiment was however unsuccessful.
- PostDetecting Changes on the ISS Autonomously with 3D Point Clouds: An Unsupervised Learning Approach Using GMM Clustering(2023) Santos, Jamie; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Chalmers University of Technology / Department of Mechanics and Maritime Sciences; Wahde, Mattias; Coltin, BrianNew space habitats, such as the Lunar Gateway and subsequent stations on Mars, will be increasingly difficult to continuously staff with astronauts, necessitating robotic support in the absence of humans. In this work, an unsupervised change detection algorithm requiring only 3D depth data as input is proposed in order to enable autonomous robotic caretaking. Scene change detection is necessary for robotic caretakers to accomplish many routine tasks, such as map updating and surveillance of the habitat. Upon the International Space Station (ISS), Astrobee is one such robot used for development and demonstration of these technologies under the Integrated System for Adaptive Autonomous Caretaking (ISAAC) project. Current scene analysis developed within the ISAAC project uses semantic localization to detect manually labeled objects within the map. However, this is not a sustainable approach for generalized change detection, as thousands of images must be captured and labeled for accurate results. In contrast, the proposed algorithm uses an unsupervised GMM clustering algorithm to compare “before” and “after” point clouds of the scene, and is therefore capable of detecting changes in the scene without being restricted to manually labeled objects. Experiments with data collected at NASA Ames’ Granite Lab, a mock-up of the ISS, successfully demonstrate the detection of one or more object appearances or disappearances in the scene with an initial average F1 score of 74% for volumetric reconstructed maps of the scene with two and three changed objects, and 70% for the comparison of single frames from the depth camera with one object. With a number of optimizations that can be made to improve the accuracy, the source code is released to the public to promote further research and development.
- PostEfficient deep learning in space: Knowledge distillation and optimization of resource usage in a satellite(2022) Asaad, Ebaa; Larsson, Sara; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Benderius, Ola; Benderius, Ola; Anlind, Alice; von Essen, Hannes; Tranheden, WilhelmThe development of micro-satellites and machine learning (ML) has increased drastically in recent years, which has unlocked new possibilities in the field of Earth observation. One of the applications is the tracking of maritime vessels since the current tracking systems such as the automatic identification system (AIS) can be eluded by simply switching it off. This project, therefore, investigates the possibilities of applying ML in a satellite, specifically with the aim of detecting maritime vessels. An object detector (YOLOv5) was chosen for testing due to its speed, small size, and its user-friendly framework. For comparison with a simpler model, the classification models ShuffleNetV1 and a custom-built CNN model were chosen. Thereafter, for the purpose of optimization, knowledge distillation, as well as different methods for reducing resource usage, were tested. The results show that it is feasible to implement ML on board a satellite to detect maritime vessels, where the best result for YOLOv5 was 2.1 min per 10,000×10,000 pixels RGB image on the target hardware using the GPU. Using the CPU with multiple threads achieved a result of 2.2 min for the same image. Increasing the batch size did not yield better results. ShuffleNetV1 was not supported by the TFLite framework, due to a network structure called group convolutions. Neither was quantization supported by the target hardware, but it did decrease the file size of the model by half. Using knowledge distillation showed great results for the classifiers. Using ShuffleNetV1 to train the simpler CNN model yielded an increase of 12 % in accuracy. It also shows that it is possible to apply a non-supported network on the target device by distilling the knowledge to a supported network. Distilling knowledge using YOLOv5 was more difficult, due to the complexity of the network and the task of object detection. Two methods were therefore tested: using the teacher’s output (logits) as a target and using the feature maps within the network as targets (feature imitation). However, only minimal improvements were reflected in the results.
- PostEfficient neuroevolution through accumulation of experience: Growing networks using function preserving mutations(2019) Löf, Axel; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Chalmers University of Technology / Department of Mechanics and Maritime SciencesIn deep supervised learning the structure of the artificial neural network determines how well and how fast it can be trained. This thesis uses evolutionary algorithms to optimize the structure of artificial neural networks. Specifically, the focus of this thesis is to develop strategies for efficient neuroevolution. The neuroevolutionary method presented in this report builds structures through architechtural morphisms that, approximately, preserve the functionality of the networks. The intended outcome of basing the mutations on the idea of function preservation was that new architechtures would start out in a high performance parameter space region. By skipping regions of low performance, the training of previous generations can be accumulated. The proposed method was evaluated relative to version in which the preservating property of the mutations was removed. In the ablated version the parameters associated with the new structural change were randomly initialized. The two versions were benchmarked on five different regression problems. On the three most difficult problems the ablated version demonstrated better performance than the preservering version, while similar performance was observed for the two other problems. The performance difference between the two versions was inferred to a more frequent tendency for the function preserving version to get entrapped in stationary regions, compared to the ablated version. The parameter initializations associated with the ablated version allow the backpropagation to more easily escape these stationary regions. The main contribution of this work is the conclusion that in order to efficiently utilize function preserving transformations to build structures in neuroevolution there need to be some mechanism that allows the backpropagation to esacpe stationary regions. The method is expected to improve by perturbating the parameters of the networks in a way that increase the gradient.
- PostFast Rainfall runoff simulation and parameter optimization(2020) Hesslow, Daniel; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Nilsson, HåkanSince the number of flooding events are expected to rise in the coming years as a consequence of global warming, accurate simulation of such events are now more important than ever. Running simulations and seeing the effects of different possible actions serves as a very important tool to mitigate the consequences of such events. For the results of the simulations to be accurate it is important that both the parameters that govern the surface flow and the subsurface flow are known, or that they can be accurately estimated. While it is feasible to measure some parameters to sufficient accuracy, such as the topology, this is not true for for all parameters. The subsurface flow is governed by the soil characteristics at all points in the simulation space and may vary over the depth. Additionally, measuring the soil characteristics at any one point is expensive. It is, therefore, not feasible to measure the soil characteristics at all points and all depths to a sufficient accuracy. The traditional approach is to have an expert estimate all such parameters, however this is costly and if ground truth data from previous flooding events are available the parameters can instead be tuned to fit with the previous events. In this thesis an efficient and numerically accurate way to calculate the infiltration of the multi-layer Green-Ampt model is presented as well as a method for automatically optimizing the parameters of large-scale fluid simulations. The developed methods are implemented in the VISDOM-application developed by VRVIS and evaluated on different scenarios.
- PostFruit Recognition by Hierarchical Temporal Memory(2011) Mattsson, Olov; Chalmers tekniska högskola / Institutionen för tillämpad mekanik; Chalmers University of Technology / Department of Applied MechanicsFood is a requirement for living, and traded in enormous amounts everyday. The globalization has led to optimization of the supermarkets and that a lot of stores have introduced self-scanning systems at check out and payment. When articles such as vegetables and fruits are traded the process becomes slower because the packages usually do not wear barcodes, which have to be added manually. This is a problem and the purpose with self-scanning drops out. In this thesis, a recognition system is built with the purpose to be used in self-scanning systems. The system thresholds the original image into a binary image. The binary image is sent to an advanced type of Neural Network called Hierarchical Temporal Memory. Such a network is independent of color, size, spatial space and rotations. These properties make it suitable for the given task. Two sorts of fruits were tested and the algorithm gave the accurate prediction in 97.5 % when tested on previously unseen images.
- PostGAN-based water droplet removal(2022) Sophonpattanakit, Jiraporn; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Chalmers University of Technology / Department of Mechanics and Maritime Sciences; Benderius, Ola; Benderius, OlaThis master thesis project studies how generative adversarial networks (GANs) can perform raindrop removal and reconstruct the scenes in single images. The models were trained by using synthetic droplet datasets, and real droplet datasets. The real droplet datasets were from Qian et al. datasets and the Reeds datasets [1]. The synthetic droplet datasets were generated from ground-truth images from prior datasets. The process was done by using OpenGL. Then the generated images were evaluated by full-reference quality metrics and non-reference quality metrics, such as SSIM and BRISQUE, then tested in object detection by a pre-trained DETR model and evaluated by mean average precision (mAP). After comparing the quantitative quality of the images generated by models trained by real droplet datasets and synthetic datasets, the result showed that the real droplet datasets yield better image quality than the synthetic datasets. In the object detection task, though it can enhance image quality in comparison to the degraded images, the generated images did not improve the result in this aspect. Thus, it was concluded that the synthetic datasets need to be more realistic to be able to reach comparable results as the real droplets. In the object detection task, GANs generated images from the information in the latent space. As a result, there were some objects which were corrupted, and this made the object detection model miss classify the objects. Apparently, it may not be suitable for high precision and safety tasks. In the aspect of the automated evaluation system, this thesis project investigated the design pattern for the conceptual design of the system. However, further functional and non-functional requirements are needed to be clarified for future implementa tio
- PostHuman intent-recognition system for safety-critical human-machine interactions(2020) Künzler, Simon; Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper; Boyraz Baykas, Pinar; Boyraz Baykas, PinarThe aim of this thesis was to investigate the potential of eye tracking technology, to help recognizing the intent of humans when working with a machine under shared control. An experiment was designed to study the eye gaze behaviour of test subjects, while manipulating a two degrees-of-freedom (DOF) SCARA robot. The subjects were given the task to maneuver the end-effector of the robot through a sequence of LEDs located on the robot action plane. The LED sequence was different for each experiment run and not known by the subjects before the start of each run. In the first step, eye gaze data was collected while the robot was unactuated. The fixation point of the subjects gaze was 4.5 times more likely to be in the proximity of the goal LEDs they intended to connect, opposed to fixating a point outside of the intended area. In addition, when the subjects planned to move from one LED to the next, the subject’s gaze tended to fixate on the next LED between one and two seconds before reaching the position with the robot end-effector, depending on how much distance the subject had to cover when moving from the current LED to the next. After reaching the fixated position, the gaze is shifted almost immediately (with 0.1-0.2s delay) onto the next LED, while movement onset is delayed about 0.5 seconds. This information was then used to develop an algorithm to predict which LED a subject is intending to reach. While performing a second set of tests, more data was collected, but this time under shared-control with the robot. The implemented algorithm was able to successfully identify the next goal LED in the subject’s planned path and to provide assistance in the movement of the robot arm. How far ahead of time the goals were recognized was dependent on how soon the subjects gaze shifted from a reached LED to their next planned goal LED. If the subject fixates on a goal LED 0.3s before initiating the movement towards it, the robot was able to perform the whole movement between the LEDs. In most cases the algorithm initiated the support half-way through the planned motion of the subjects. No significant differences in the subjects gaze data between passive robot manipulation and shared control could be identified.