Internships & AI thesis projects

AboutFootball event datasets sometimes contain missing or incomplete actions, even when those actions are visible in the match footage. This project investigates whether computer vision and sequential deep learning models can combine video evidence with surrounding match events to identify and reconstruct those missing actions. Students will compare different modelling approaches and evaluate how reliably the system distinguishes genuine omissions from actions that never occurred.

OutcomeA benchmark and a comparison of video-only, event-only and combined methods.

References: Ochin et al. (2026), FOOTPASS · Shin et al. (2025), Soccer-CLIP · Kim et al. (2026), ELASTIC · Denize et al. (2024), COMEDIAN

AboutTraining computer vision models typically requires large amounts of manually annotated footage, which is expensive and time-consuming to produce. This project investigates whether adapting pretrained video models to football and intelligently selecting training examples can reduce annotation requirements. Students will experiment with different representations and active-learning strategies to determine which combinations achieve strong performance using fewer labelled examples.

OutcomeLearning curves showing which sampling approaches save annotation effort.

References: Yang et al. (2026), SoccerMaster · Shin et al. (2025), Soccer-CLIP · Tong et al. (2022), VideoMAE · Sener and Savarese (2018), Core-Set Active Learning · Ash et al. (2020), Deep Batch Active Learning

AboutKnowing where players are on the pitch is the basis for most football analysis, but footage from a single camera only shows players in image coordinates. This project investigates how pitch registration can map players from one camera view to real pitch positions, and how a model can estimate how reliable each of those positions is. Students will compare registration and localisation approaches and evaluate both positional accuracy and the quality of the uncertainty estimates.

OutcomeA pitch-position benchmark and trustworthy spatial uncertainty estimates.

References: Gutiérrez-Pérez and Agudo (2026), PnLCalib · Ziegler et al. (2026), AuxFlow · Cuevas et al. (2026), Viewpoint-Invariant Soccer Pitch Registration · Somers et al. (2024), SoccerNet Game State Reconstruction

AboutTracking football players across a match is difficult because cameras move, players overlap, and individuals frequently disappear from view. Most tracking systems operate primarily in image coordinates, which can make it difficult to distinguish camera movement from actual player movement. This project investigates whether incorporating real pitch coordinates into the tracking process can improve player identity continuity and recovery after occlusion.

OutcomeA tracking comparison covering identity switches, recovery and registration error.

References: Chen et al. (2025), FieldMOT · Gao et al. (2025), Multiple Object Tracking as ID Prediction · Scott et al. (2024), TeamTrack · Zhang et al. (2022), ByteTrack · Somers et al. (2024), SoccerNet Game State Reconstruction

AboutFootball analysts often describe players through their positions, movements, or tactical roles rather than simply using their names. This project investigates whether vision-language models can connect natural-language descriptions, such as “the winger making an overlapping run,” to the correct player trajectory in match footage. Students will develop and evaluate methods combining language, visual appearance, movement patterns, and spatial relationships.

OutcomeA referring-player dataset and language-to-track evaluation.

References: Wu et al. (2023), Referring Multi-Object Tracking · Li et al. (2025), Language Decoupling with Fine-Grained Knowledge Guidance · Chamiti et al. (2025), ReferGPT · Radford et al. (2021), CLIP

AboutAutomatically identifying football players is challenging because teammates wear similar clothing and jersey numbers are not always visible. This project investigates how multiple sources of information, including jersey recognition, visual appearance, team membership, and spatial context, can be combined to improve identification. Students will evaluate different evidence-fusion approaches and investigate how models can express uncertainty instead of making incorrect identity assignments.

OutcomeAn identity benchmark that also measures when the model should answer “unknown”.

References: Koshkina and Elder (2024), Jersey Number Recognition in Sports Video · Grad (2025), Uncertainty-Aware Jersey Number Recognition · Somers et al. (2024), SoccerNet Game State Reconstruction · Yang et al. (2026), SoccerMaster

AboutForward Football uses its Situational Performance Framework (SPF) to evaluate how players contribute to tactical situations and the quality of their individual actions. This project investigates whether deep learning models can learn from expert assessments and structured match evidence to estimate tactical contribution and four distinct performance dimensions: Game Understanding, Decision Making, Situational Speed, and Situational Accuracy. Students will compare modelling approaches while investigating prediction reliability, interpretability, and agreement with expert evaluations.

OutcomeAn evaluated model for Tactical Contribution and distinct GA / DM / SiS / SiA ratings, with uncertainty.

References: Fernández et al. (2021), Expected Value of Soccer Possessions · Nakahara et al. (2023), Action Valuation with Multi-Agent Deep RL · Overmeer et al. (2025), Revisiting Expected Possession Value · Wang et al. (2024), TacticAI

AboutFootball players continuously make decisions whose quality depends on the surrounding game situation, available options, and potential outcomes. This project investigates whether offline deep reinforcement learning can learn meaningful action values from historical match events, possession sequences, and reconstructed football situations. Students will compare established value-learning methods and reward formulations while examining the limitations of evaluating alternative decisions using observational data.

OutcomeA conservative action-value benchmark; no claim that an unplayed alternative would have succeeded.

References: Rahimian et al. (2024), In-Game Soccer Outcome Prediction with Offline RL · Nakahara et al. (2023), Action Valuation with Multi-Agent Deep RL · Ide et al. (2026), Expandable Decision-Making States · Pace et al. (2025), Preference Elicitation for Offline RL · Kausik et al. (2024), Offline Policy Evaluation under Confounding

AboutUnderstanding a football situation requires more than observing an isolated action; it also involves the preceding sequence of events, player positioning, and interactions between teams. This project investigates how temporal deep learning and spatial representation models can combine these different information sources into meaningful game-state representations. Students will compare event-only models with sequence-based and computer-vision-enhanced approaches to measure the additional value of each data source.

OutcomeA reproducible comparison of event-only, history-aware and CV-enriched representations.

References: Yeung et al. (2025), OpenSTARLab · Ochin et al. (2026), FOOTPASS · Kim et al. (2026), ELASTIC · Fernández et al. (2021), Expected Value of Soccer Possessions · Wang et al. (2024), TacticAI

AboutForward Football works with multiple sources of football information, including match events, video, player records, annotations, and computer vision outputs. This project focuses on connecting, organising, and validating these datasets so that they can support reliable analysis and machine learning development. Students will work with data engineering and analytical methods to create reproducible datasets, identify data-quality issues, and generate useful football insights.

OutcomeLinked data views, quality report, versioned dataset candidates and useful insights.

References: Gebru et al. (2021), Datasheets for Datasets · Breck et al. (2017), The ML Test Score · Bassek et al. (2025), Integrated Dataset of Spatiotemporal and Event Data in Elite Soccer · Yeung et al. (2025), OpenSTARLab

AboutComputer vision models often perform differently across football matches because of changes in camera angles, lighting, player appearance, and image quality. This project allows students to investigate and improve a specific football vision task, such as player detection, ball detection, participant classification, or pitch-related visual recognition. Students will prepare annotated datasets, experiment with model adaptation or fine-tuning, and evaluate improvements against established baselines.

OutcomeA controlled model comparison, error analysis and reproducible candidate.

References: Deliège et al. (2021), SoccerNet-v2 · Yang et al. (2026), SoccerMaster · Somers et al. (2024), SoccerNet Game State Reconstruction · Scott et al. (2024), TeamTrack

AboutRecognising the same football player across different frames or video segments is challenging because appearance changes with camera perspective, resolution, and occlusion. This project investigates how machine learning models can generate visual representations, known as embeddings, that capture useful similarities between player observations. Students will construct evaluation datasets, compare representation-learning approaches, and develop reusable components for player retrieval and tracklet association.

OutcomeA versioned tracklet benchmark and reusable player-embedding component.

References: Aharon et al. (2022), BoT-SORT · He et al. (2023), FastReID · Scott et al. (2024), TeamTrack · Chen et al. (2025), FieldMOT

AboutFootball tactics involve relationships between player positioning, available space, team structure, and movement during different game situations. This project focuses on converting reviewed football annotations and reliable spatial information into measurable tactical features and interpretable analysis. Students will investigate selected tactical concepts, develop reproducible datasets and analytical methods, and explore how the resulting information could support football performance evaluation.

OutcomeA traceable tactical dataset and interpretable analysis or prototype.

References: Fernández et al. (2021), Expected Value of Soccer Possessions · Wang et al. (2024), TacticAI · Somers et al. (2024), SoccerNet Game State Reconstruction · Gutiérrez-Pérez and Agudo (2026), PnLCalib

AboutTool calling and retrieval over structured match data, translating natural-language questions into queries, and scoping answers by club and user role. The chatbot is integrated into Technoleon; building a standalone RAG system is out of scope.

OutcomeAn accuracy comparison on a test-question set covering correct answers, invented answers and response time.

AboutPlayer and ball detection, pitch calibration, mapping image coordinates to pitch coordinates, and 3D playback of the resulting positions. The focus is on the video design overlay and 3D viewer, not on the computer vision work itself.

OutcomeA position-accuracy evaluation against known pitch positions, together with a working 3D viewer.

AboutStructured synthetic states, graph models and counterfactual evaluation.

OutcomeReal-versus-synthetic training comparisons grounded in expert-defined principles.

AboutSelf-supervised video pretraining, transfer learning and cross-task evaluation.

OutcomeA transferable representation benchmark on more than one football task.

AboutMulti-agent Transformers, trajectory prediction and uncertainty calibration.

OutcomeForecasting results against geometry-only and simpler motion baselines.

AboutMonocular 3D vision, camera geometry and temporal/physical constraints.

OutcomeA 3D ball-reconstruction feasibility benchmark, including failure conditions.