Browsing by Author "Matiisen, Tambet, juhendaja"

Now showing 1 - 20 of 24

Comparing Output Modalities in End-to-End Driving
(Tartu Ülikool, 2022) Aidla, Romet; Matiisen, Tambet, juhendaja; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Self-driving car technology has made significant steps in the last ten years with the advancements in neural networks. The first autonomous vehicles are driving in San Francisco and Beijing. One of the promising approaches is end-to-end driving, where a neural network transforms an input image from a camera to output commands to control the vehicle. The most common output modalities are steering angle and trajectory. Both have been extensively benchmarked but not compared in similar settings. Metrics are usually calculated off-policy using a separated test dataset or on-policy using a simulator, but these have proven to correlate weakly with real-life performance. In this thesis, the comparison is made using an autonomous vehicle driving on WRC Rally Estonia tracks. The results show that the trajectory prediction approach is better at road positioning and recovering from non-ideal trajectories, which results in fewer situations where the safety driver has to take over.
Creating High-Definition Vector Maps for Autonomous Driving
(Tartu Ülikool, 2021) Sepp, Edgar; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Autonomous driving holds many promises for transportation - increased safety, lower costs, and less burden to the environment. In light of some recent accidents, it is clear that the technology is not fully ready yet, and the robustness and research in the area need to be increased. Most of the autonomous driving solutions rely on high-definition maps (HD maps) - specialized lane-level maps with very high locational accuracy. Mobile mapping cars (specially equipped vehicles with sensors for map data collection) by big mapping companies are used to collect the data for creating HD maps. Along with required data processing the creating and keeping the HD maps up to date in a changing world is very costly. Availability of the HD maps would considerably lower the bar for adopting autonomous driving at large. To the best of the author’s knowledge, there are no freely available HD maps for self-driving available for Estonia. To be able to conduct research experiments with the University of Tartu's Autonomous Driving Lab (UT ADL) self-driving platform, such maps had to be created. Several available tools for creating the maps and existing data sources were reviewed. The custom workflow was created for mapping and a tool to convert the HD vector map to Autoware vector map format was created. Finally, quantitative measures about time estimates needed to create the HD vector maps and their usage in UT ADL were given.
Creation of Digital Twin for Tartu
(Tartu Ülikool, 2024) Mitt, Allan; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Enne isejuhtiva sõiduki testimist reaalses maailmas on tõhusam katsetada seda simulatsiooni keskkonnas, mis imiteerib pärismaailma. See lõputöö kirjeldab Tartu linna digikaksiku loomist, kasutades avalikke ruumiandmeid ja rakendades erinevaid automatiseerimistehnikaid tööprotsessi kiirendamiseks. Lõpptulemus on ühilduv isejuhtivate sõdukite tarkvaraga ning tartlastele väga äratuntav.
Developing a Volleyball Game with an AI Opponent Using Reinforcement Learning
(Tartu Ülikool, 2021) Marran, Tanel; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This thesis integrates reinforcement learning into a game development project by creating a competitive volleyball game, where the user can play against an artificial intelligence (AI) trained using reinforcement learning techniques. The work elaborates on what reinforcement learning is, brings forth some of the challenges of adding machine learning to a game, describes the development environment Unity and its machine learning package ML-Agents as well as analyzes the finished game and its AI.
Energy-Based Models for End-to-End Autonomous Driving
(Tartu Ülikool, 2022) Baliesnyi, Mykyta; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Energy-based models (EBMs), a promising class of machine learning models, have shown impressive results in several domains, from natural language generation to computer vision. Learning to imitate expert demonstrations using an EBM has recently achieved state-of-the-art results in robotics, made possible by EBMs’ better ability to handle multimodal probability distributions and learn behavior with abrupt command changes. In this work, EBMs are tested for the first time in the end-to-end autonomous driving domain on a real car. As a result, it is discovered that a simple EBM variant performs slightly better and is more stable than a baseline conventional neural network architecture. At the same time, EBMs turn out to exhibit a higher variability of predictions over time, or whiteness. As a solution to this problem, this work introduces a regularization technique that makes the predictions more smooth over time. In addition, an energybased uncertainty metric is proposed, but its usefulness could not be assessed with sufficient reliability due to an insufficient number of real car evaluations. The thesis suggests several ideas for future work, such as using a different sampling method and comparing against mixture density networks.
Georeferenced Visual SLAM
(Tartu Ülikool, 2023) Mägi, Erik; Matiisen, Tambet, juhendaja; Sepp, Edgar, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This thesis presents a complementary localization solution for taxis and ride-hailing operators in situations where GNSS is unavailable or unreliable. The proposed method leverages monocular visual SLAM techniques, specifically the ORB-SLAM 3 library, to create a map of the environment and localize within it. The system uses a car-mounted camera for image capture and an advanced GNSS receiver to record accurate ground truth. This data is then used as input for training a deep learning model to transform SLAM coordinates into georeferenced coordinates. The thesis explores different approaches to solving the coordinate transformation problem, including linear transformation, machine learning regression algorithms, and deep learning with neural networks. Results show that the deep learning based approach provides the best localization accuracy, surpassing that of modern smartphone GNSS. The study contributes a practical solution for real-time localization for ride-hailing operators when GNSS is compromised, with the potential for future implementation using smartphone cameras.
In Search of the Best Activation Function
(Tartu Ülikool, 2022) Liibert, Marti Ingmar; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
The choice of an activation function in neural networks can have great consequences on the performance of the network. Designing and discovering new activation functions that increase the performance or solve problems of existing activation functions is an active research field. In this thesis, a kind of trainable activation function is proposed - a weighted linear combination of activation functions where the weights are normalized using Softmax, inspired by the DARTS network architecture search method. The activation function is applied at the layer, kernel, and neuron levels. Optimizing the activation function weights is done on training loss and validation loss, as was done in DARTS. The activation function here was tested on two simple datasets, sine wave, and spiral datasets, on image classification tasks and on a robotics task. In the case of image classification, on CIFAR10 using the trainable activation function for initial training the accuracy increased 5% over the baseline, on ImageNet the accuracy increased 1% over the baseline. For the robotics task, CartPole, the mean reward increased by 10 points out of a maximum of 200 when using the already learned activation functions in the case of Deep Q-learning. In the case of Proximal Policy Optimization, the mean reward increased by 2 points approximately over the baseline. For future work, more difficult tasks could be explored for robotics tasks and longer initial search could be explored for image classification tasks.
Inimlik kiirus kurvides
(Tartu Ülikool, 2021) Rudi, Eduard; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Bakalaureusetöö eesmärk oli leida võrrand, mille sisendiks on joone kõverus ja tagastab teelõigu kiiruse, mis on konstrueeritud inimese juhtimisandmete põhjal. Töös antakse lühiülevaade tänapäeva isejuhtimisüsteemi tarkvara ehitamise lähenemistest. Samuti tutvustatakse varasemaid uuringuid. Töös kirjeldatakse põhjalikult andmete töötlemisest ning mida nendega tehti. Kirjeldatakse detailselt lahti, kuidas võrrand implementeeriti Autoware’i ja kuidas algoritm arvutab kiirused mingi teekonnale. Lõpuks antakse ülevaade testimise tulemustest.
Lane Centerline Detection from Orthophotos using Transformer Networks
(Tartu Ülikool, 2024) Pilve, Karl-Johan; Matiisen, Tambet, juhendaja; Sepp, Edgar, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Sõiduradade andmeid sisaldav täppiskaart on isejuhtivate sõidukte opereerimisel väga oluline komponent. Samas on täppiskaardi loomine suurema ala kohta sageli ajakulukas ning töömahukas protsess. Hiljuti on transformeri arhitektuuri kasutavad tehisnärvivõrgud näidanud paljulubavaid tulemusi masinnägemise valdkonnas. Üks selline näide on RNGDet, mis genereerib iteratiivselt teedevõrgu graafi aerofotode põhjal. Käesolevas töös uuritakse võimalust peenhäälestada RNGDet mudelit sõiduradade andmetega, et genereerida kogu Tartu linna kattev sõiduradade graaf kasutades kõrge resolutsiooniga ortofotosid. Töös saadud tulemused näitavad, et RNGDeti on põhimõtteliselt võimalik kasutada sõiduradade graafi genereerimiseks. Samas oleks mudeli arhitektuuris vaja tõenäoliselt teha suuremaid muudatusi, et võtta arvesse teedevõrgu ja sõiduradade andmete vahelisi erinevusi. Kuna ortofotodel ei pruugi alati olla kogu vajaliku informatsioon sõiduradade õigesti genereerimiseks, siis kõige paremaid tulemusi andis mudel, mis kasutas ortofotodele rasterdatud mõõdistussõiduki poolt kogutud töötlemata GPS trajektoore. Saadud tulemused näitavad veel, et täppiskaardi jaoks sobiva kvaliteediga sõiduradade genereerimiseks oleks vaja koguda täiendavaid treeningandmeid.
Learning Competitive Minecraft Minigames with Reinforcement Learning
(Tartu Ülikool, 2022) Sisask, Laur; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
In recent years, deep reinforcement learning methods have successfully been used to play complex games like Go, StarCraft II, and Dota 2 at a professional level. In this thesis, reinforcement learning methods are used to train artificial agents in the game of Minecraft. Various competitive 1v1 Minecraft minigames from one of the most popular Minecraft servers Hypixel are selected. Deep neural networks are trained to play each of these games using proximal policy optimization algorithms and self-play. In all the games, artificial agents were able to play the game at least on a beginner level. In one game, the agent reached the level of expert human players.
Navigatsioonirakenduse sisendi renderdamine autot juhtivale närvivõrgule
(Tartu Ülikool, 2021) Maks, Erik Marcus; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This thesis describes the making of a navigation application that can be used as an input for a neural network. The application will receive GPS coordinates and will render step-by-step directions on screen. This rendering will be used together with camera images as an input for the neural network.
Object Recognition Using a Sparse 3D Camera Point Cloud
(Tartu Ülikool, 2023) Tiirats, Timo; Matiisen, Tambet, juhendaja; Bogdanov, Jan, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
The demand for higher precision and speed of computer vision models is increasing in autonomous driving, robotics, smart city and numerous other applications. In that context, machine learning is gaining increasing attention as it enables a more comprehensive understanding of the environment. More reliable and accurate imaging sensors are needed to maximise the performance of machine learning models. One example of a new sensor is LightCode Photonics’ 3D camera. The thesis presents a study to evaluate the performance of machine learning-based object recognition in an urban environment using a relatively low spatial resolution 3D camera. As the angular resolution of the camera is smaller than in commonly used 3D imaging sensors, using the camera output with already published object recognition models makes the thesis unique and valuable for the company, providing feedback for LightCode Photonics’ current camera specifications for machine learning tasks. Furthermore, the knowledge and materials could be used to develop the company’s object recognition pipeline. During the thesis, a new dataset is generated in CARLA Simulator and annotated, representing the 3D camera in a smart city application. Changes to CARLA Simulator source code were implemented to represent the actual camera closely. The thesis is finished with experiments where PointNet semantic segmentation and PointPillars object detection models are applied to the generated dataset. The generated dataset contained 4599 frames, of which 2816 were decided to use in this thesis. PointNet model applied to the dataset could predict the semantically segmented scene with similar accuracy as in the original paper. A mean accuracy of 44.15% was achieved with PointNet model. On the other hand, PointPillars model was unable to perform on the new dataset.
Põllukultuuride tuvastamise masinõppe mudeli tunnuste olulisuse hindamine
(Tartu Ülikool, 2021) Järveoja, Mihkel; Voormansik, Kaupo, juhendaja; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Remotely sensed, in particular satellite data, is already widely used in agricultural parcels monitoring, and this trend is not showing signs of diminishing. Wide range of machine learning algorithms have significantly reduced the burden to interpret bulky and often complex satellite data, contributing to the exploration of new use-cases and services. In this study Random Forest classification model is used to separate 28 crop type classes in Estonia. Input data consisted of two seasons (2018, 2019) of Estonian agricultural parcels and features calculated from Sentinel-1 and Sentinel-2 satellite images, meteorological records and soil maps. Achieved multiclass weighted F1 score for year 2018 test set was 0.82 and for year 2019 0.85. Among most important features were Sentinel-1 VH and VV polarization back-scatter intensities and Sentinel-2 PSRI, NDVI and TC-vegetation indices. It was discovered that Sentinel-2 features were more prominent in early (May) and late season (August), but during mid-season (June, July) their importance decreased significantly. Sentinel-1 back-scatter features were more important during mid-season. It was concluded, that using both radar and optical satellite data ensure better classification result than using any of them separately, since they complement each other.
Real-time 3D Object Detection on Point Clouds
(Tartu Ülikool, 2020) Ozipek, Enes; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Abstract: The demand for precise and fast object detection frameworks has increased since the autonomous vehicle industry started to attract more attention. While the progress made so far in 2D object detection task with state-of-the-art approaches such as convolutional neural networks seems promising, we still struggle to obtain the same level of performance in 3D modalities such as lidar point clouds. The main reasons are that point cloud is sparse and in 3D while state-of-the-art 2D object detection models work on camera images. Some of the early works have tried to ease the aforementioned challenges using either 3D convolutional neural networks or bird’s eye view approaches, nevertheless, they were not able to achieve the desired level of performance in 3D perception. PointPillars is one of the recent models running fast with a good accuracy on point clouds. Its main advantage arises from the way it encodes the points in pillars into spatial features using PointNet. It basically divides the whole point cloud into grids of vertical pillars and applies state-of-the-art 2D detection network on this top-down view in which spatial features are encoded. Even though this operation enables the network to keep the positional information of the points within each pillar, yet, it does not take into account the point densities in different parts of the point cloud. This thesis aims to improve PointPillars network by utilizing the positional encoding and extending the detection area. Positional encoding helps the network utilize positional features by introducing two additional input channels before each convolutional and deconvolutional layer. Additionally, different positional encoding schemes are compared to have more insight about the effectiveness of the positional channels introduced. Moreover, this thesis also presents a simple scheme to train 360-degrees model with ground truths provided for only camera Field-of-View (FOV). Positional encoding scheme provides better accuracy at a similar speed as the original network. On the other hand, even though 360-degrees model is supposedly the type of a model that should be used with lidar, in experiments, it is observed that it outputs many False-Positives (FPs).
Replicating DeepMind StarCraft II reinforcement learning benchmark with actor-critic methods
(2018) Ring, Roman; Kuzovkin, Ilya, juhendaja; Matiisen, Tambet, juhendaja; Tartu Ülikool. Matemaatika ja statistika instituut; Tartu Ülikool. Loodus- ja täppisteaduste valdkond
Reinforcement Learning (RL) is a subfield of Artificial Intelligence (AI) that deals with agents navigating in an environment with the goal of maximizing total reward. Games are good environments to test RL algorithms as they have simple rules and clear reward signals. Theoretical part of this thesis explores some of the popular classical and modern RL approaches, which include the use of Artificial Neural Network (ANN) as a function approximator inside AI agent. In practical part of the thesis we implement Advantage Actor-Critic RL algorithm and replicate ANN based agent described in [Vinyals et al., 2017]. We reproduce the state-of-the-art results in a modern video game StarCraft II, a game that is considered the next milestone in AI after the fall of chess and Go.
Sample-efficient Online Learning in a Physical Environment
(Tartu Ülikool, 2020) Liivak, Martin; Matiisen, Tambet, juhendaja; Paat, Rainer, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Autonomous driving has been seen as the next breakthrough in transportation. Autonomous vehicles employ a variety of sensors to understand their surroundings, for example multiple cameras, ultrasound sensors, and LiDARs. In this work, a much smaller scale radio-controlled cars, that only carry a central camera, are used. Their effectiveness as a test-bed for validating autonomous driving methods is evaluated. Multiple neural network architectures were proposed, among which a convolutional neural network was selected as the best candidate. The network was then trained using both supervised learning and online learning, the results of which were then compared. Experiments show that online learning in a physical environment, while costly, is a significant improvement over pure supervised learning. Additionally the radio-controlled cars proved to be a good comparative test-bed for evaluating model performance in an interactive physical environment.
Semi-automatic generation of Lanelet2 maps for autonomous driving
(Tartu Ülikool, 2023) Sagris, Valentina; Sepp, Edgar, juhendaja; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
High-definition (HD) maps are essential in autonomous driving (AD) by providing data about the static driving environment for vehicle localisation, route planning, perception and manoeuvre decision-making. To conduct research in autonomous driving, the Autonomous Driving Lab (ADL) of the University of Tartu needs HD maps. This thesis presents the database structure and the effective workflow for the semi-automatic generation of HD map elements. In particular, it tests the capability of the PostGIS spatial database in helping to minimise manual work. The main goals of this thesis can be summarised into the following topics. The ADL's previous mapping efforts were overviewed, and the unified database design for storing spatial data required in various HD map formats was proposed. Then previously collected data was used to develop an algorithm for the semi-automatic generation of missing elements – spatial features needed to construct proper lanelets, the fundamental primitives of maps in Lanelet2 format. Further on, the Lanelet2 requirements for the shared bounds and auto-mated finding of semantic relationships between primitives that constitute a lanelet were addressed. The critical step here was producing spatial elements, which we call 'relations' that establish the spatial relationships between a centreline and its bounds in the database. Finally, the data converter for HD maps in Lanelet2 format was proposed. It transforms spatial primitives of PostGIS into primitives of Lanelet2, which are nodes, ways and relations. Due to PostgreSQL's capability to store the XML datatype, elements for each primitive were created in the database. Further on, the database was accessed from the python script, where XML root was made, and the following elements were loaded from the database into the root to create a proper Lanelet2 file. The workflow was tested on two sites: Lai-Jakobi-Kroonuaia and Narva-Roosi-Puiestee. It helped to assess the algorithm's robustness in various road network configurations and improve its performance. It turned out that the algorithm performance is very good with standard road structures where plain street stretches meet in T-shape or X-shape intersections. Also, the algorithm is capable of performing well in areas of complex street intersections, but some human assistance is needed. The statistics obtained for comparison of the proposed solution with a fully manual process of map elements digitisation demonstrated a considerable reduction in time and human efforts in HD map creation. Due to the PostGIS functionality, the data geoprocessing was impres-sively fast. Depending on road structure complexity, the amount of effort needed per one kilometre of lanes can drop by 45% to 65%.
Sim-to-Real Generalization of Computer Vision with Domain Adaptation, Style Randomization, and Multi-Task Learning
(Tartu Ülikool, 2020) Liik, Hannes; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
In recent years, supervised deep learning has been very successful in computer vision applications. This success comes at the cost of a large amount of labeled data required to train artificial neural networks. However, manual labeling can be very expensive. Semantic segmentation, the task of pixel-wise classification of images, requires painstaking pixel-level annotation. The particular difficulty of manual labeling for semantic segmentation motivates research into alternatives. One solution is to use simulations, which can generate semantic segmentation ground truth automatically. Unfortunately, in practice, simulation-trained models have been shown to generalize poorly to the real world. This work considers a simulation environment, used to train models for semantic segmentation, and real-world environments to evaluate their generalization. Three different approaches are studied to improve generalization from simulation to reality. Firstly, using a generative image-to-image model to make the simulation look realistic. Secondly, using style randomization, a form of data augmentation using style transfer, to make the model more robust to change in visual style. Thirdly, using depth estimation as an auxiliary task to enable learning of geometry. Our results show that the first method, image-to-image translation, improves performance on environments similar to the simulation. By applying style randomization, the trained models generalized better to completely new environments. The additional depth estimation task did not improve performance, except by a small amount when combined with style randomization.
Simulation of Tartu City Centre for Testing Autonomous Vehicles
(Tartu Ülikool, 2021) Mitt, Allan; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Before testing an autonomous vehicle in the real world, it would be more efficient to try it in the simulation representing the real-life setting. This thesis aims to describe the designing process behind creating an adequate representation of Tartu city centre within the simulation.
Simulations for Training Machine Learning Models for Autonomous Vehicles
(Tartu Ülikool, 2020) Toompea, Kertu; Tunnel, Raimond-Hendrik, juhendaja; Matiisen, Tambet, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Training machine learning models for autonomous vehicles requires a lot of data which is time consuming and tedious to label manually. Simulated virtual environments help to automate this process. In this work these virtual environments are called simulations. The goal of this thesis is to survey the most suitable simulations for off-road vehicles (while not discarding the urban option). Only the simulations which provide labeled output data, are included in this work. The chosen 12 simulations are surveyed based on the information found online. The simulations are then analyzed based on the predefined features and categorized according to their suitability for training machine learning models for off-road vehicles. The results are shown in a table for comparison. The main purpose of this work is to map the seemingly large landscape of simulations and give a compact picture of the situation.