Teaching

Accurate estimation of traffic flows is fundamental for intelligent transportation systems, urban mobility management, and sustainable planning. Traditionally, such estimations rely on extensive sensor infrastructures, including inductive loops, GPS trajectories, or mobile phone data. However, in many cities – particularly small and medium-sized ones – camera counts at intersections remain the most available and cost-effective data source. These counts provide only partial and noisy observations of the underlying traffic dynamics, making the problem of network-wide flow estimation challenging. Recent advances in Graph Neural Networks (GNNs) offer powerful tools for modeling spatiotemporal processes on road networks, learning flow patterns directly from graph-structured data. At the same time, agent-based models (ABMs), though computationally demanding, provide interpretable simulations that capture the microscopic decision-making of vehicles and drivers, such as route choice and turning movements. This project proposes a hybrid approach that integrates GNN learning with ABM insights to leverage the strengths of both paradigms.

Autonomous vehicles (AVs) rely heavily on perception systems to ensure safe navigation in complex urban environments. While recent advances in multimodal large language models (MLLMs) have demonstrated strong capabilities in scene understanding, reasoning, and providing natural-language explanations, their application to real-time prediction of dangerous or risk-critical situations remains underexplored. Most existing approaches either (i) analyze static frames or short clips offline, (ii) generate risky scenarios for testing, or (iii) provide risk recognition without strict latency guarantees. However, urban driving demands continuous, low-latency anticipation of hazards – such as a pedestrian suddenly entering the road or a vehicle running a red light – where milliseconds of delay may determine safety outcomes. This thesis aims to close this gap by integrating temporal dynamics modeling into multimodal LLM frameworks, enabling predictive risk recognition in real time.

Object detection in city environments is critical for applications such as autonomous driving, intelligent surveillance, and urban planning. As the complexity and scale of these applications grow, so does the demand for computational resources, leading to increased energy consumption. Traditional Transformer-based models, while powerful, are often energy-intensive. On the other hand, Spiking Neural Networks (SNNs), inspired by biological neural processes, promise significant energy savings due to their event-driven nature. This project explores the potential of SNNs as a more energy-efficient alternative to Transformers. The aim is to analyze the comparative performance of Spiking Neural Networks (SNNs) and Transformer-based architectures in object detection within urban scenes. The energy efficiency analysis focuses on determining which neural network paradigm offers superior performance from both (accuracy and liability) while minimizing power consumption. By leveraging neuromorphic hardware and optimizing Transformer models, the research seeks to contribute to developing sustainable and efficient AI systems for real-time urban environment analysis.

Accurate forecasting allows operators to strategically reposition bikes, ensuring availability where and when it’s most needed. This directly contributes to a better user experience by reducing the frustration of finding empty docks or unavailable bikes. Furthermore, effective demand prediction is crucial for the equitable and sustainable expansion of bike-sharing systems, helping to mitigate urban congestion and promoting environmentally friendly transportation. What You’ll Build? Delve into the world of mobility modelling and machine learning to decide which is the best model for this task. Implement and validate a machine learning model to forecast demand in existing bikeshare stations using historic data and other information, such as weather and time of day.

Location data is the backbone of today’s shared mobility services—but GPS alone is vulnerable to spoofing, tampering, and errors. How can we guarantee that an e-scooter was really at a parking bay, at the time a ride started or ended? This project transforms everyday e-scooters into mobile witnesses in a secure urban mesh. Using Raspberry Pi pods, UWB ranging, and BLE Direction Finding, scooters and curbside anchors will generate cryptographically verifiable proofs of location (PoL)—without touching scooter firmware. What You’ll Build? Rugged Raspberry Pi–powered PoL pods mounted on scooters, anchor stations with GNSS time sync and multi-antenna BLE/UWB arrays, a PoL protocol implementation with quorum attestations and transparency logging, and a live field pilot: ride start/end proofs, parking compliance, and anti-theft tracking. This is a hands-on thesis project: you will design hardware, implement secure protocols, and test your system on the streets. Perfect for those passionate about IoT security, smart mobility, and applied cryptography.

Multi-object tracking (MOT) is a fundamental task in the field of computer vision that involves tracking multiple objects across video frames. It extends beyond simple object detection by maintaining consistent identities for each object throughout a sequence, despite challenges such as occlusion, abrupt movements, and changes in appearance. This capability is crucial for applications requiring detailed analysis of object interactions and movements over time. On the other hand, Mamba is a cutting-edge deep learning architecture designed for efficient sequence modeling, particularly addressing the limitations of traditional Transformer models in handling long sequences. This project aims to revolutionize multi-object tracking (MOT) by integrating the Mamba architecture as a long-sequence memory model. The goal is to enhance the ability of MOT systems to handle complex motion patterns and occlusions, which are common challenges in dynamic environments such as video surveillance and autonomous vehicles.

Deep neural networks (DNNs) have succeeded remarkably in various AI applications, including computer vision, speech recognition, and machine translation. However, their high computational demands result in significant energy consumption, which raises operating costs for data centers and limits their deployment on mobile devices with constrained energy budgets. As local processing on mobile devices gains popularity due to privacy, security, and latency concerns, designing energy-efficient DNNs becomes essential for advancing mobile AI applications. Hence, in this thesis, we will focus on designing a tool or technique for measuring the energy consumption of the DL model.

Tartu Smart Bike (Tartu rattaringlus) is a bike-sharing system in Tartu, providing residents with an affordable and environmentally friendly travel option. Over its five years of operation, it has accumulated a significant amount of data, including trip-level and detailed GPS data. This data can aid researchers and policymakers in designing more bikeable cities. While parts of this data have been utilized in various research projects, there is a lack of a comprehensive tool that would offer an overview of the data spanning the entire period of the system’s operation. The aim of this thesis is to create a web-based dashboard to visualize and explore Tartu Smart Bike data. The dashboard should enable users to examine bike-sharing usage from both spatial and temporal perspectives and compare different time periods. Depending on the scope, a more in-depth analysis of GPS trajectories could be included. The outcome of this work should be a web application that facilitates user-friendly exploration of the Tartu Smart Bike system.

Simulation and optimization of control strategies for ramp merging management.

In order to evaluate the performance and feasibility of a new powertrain, driving cycles has been widely applied to represent real driving conditions. Since the driving behavior varies with the local conditions, typical driving cycles are needed considering local conditions, such as weather, terrain, and traffic status. Based on the GPS data, a typical driving cycle for buses in Tartu will be developed. This thesis topic is in collaboration with ITS Lab, Tartu City and OsloMet.

Based on the GPS data of buses, the energy needed of each bus route under different weather can be derived. Therefore, the feasibility of different electric bus technologies should be analyzed. Considering the bus schedule and route-level energy consumption, optimal charging infrastructure placement and charging strategies should be developed considering battery life and lifecycle cost. This thesis topic is in collaboration with ITS Lab, Tartu City and OsloMet.

Discovering the share of different means of transport is an important component for developing sustainable transport system. In this thesis, the initial modal split model will be created using some stochastic data. Then, using online data (such as traffic count), the model will be calibrated on short time intervals. This thesis topic is in collaboration with ITS Lab and Tartu City.

Advancing in the direction of reducing the time complexity of masking the moving objects Visual SLAM input. It was proven in previous research MaskRCNN is accurate but can hardly achieve 10FPS. Objective is to employ state estimation techniques to track the moving objects and update the masking information faster then actual detection rate.

Simulator of Urban Mobility (SUMO) is an open source, highly portable, microscopic and continuous multi-modal traffic simulation package designed to handle large networks. Current implementation performs all the road network related routines (simulation, calibration etc.) using single thread hence under-utilizing the multi-core CPU. The objective is to achieve better CPU utilization by allowing multiple concurrent threads within one simulation.