Diploma Thesis: Trajectory Prediction of Vulnerable Road Users at Smart Intersections for Road Safety and Autonomous Driving Applications
Can multimodal context—motion, distance, traffic-light state, and spatial grids—improve VRU trajectory prediction over historical trajectory alone?
Project-specific research
Forecasting methods and research questions
Forecast trajectories and examine uncertainty.
The diploma thesis compares recurrent, attention-based, graph-based and probabilistic sequence models for predicting vulnerable road-user motion at smart intersections.
01 / 04
Observed trajectories + context
02 / 04
LSTM / GRU / Transformer / GNN
03 / 04
Deterministic + MDN forecasts
04 / 04
ADE / FDE / uncertainty
Research question: How does contextual information change prediction accuracy and uncertainty?
Research objective
Evaluate whether multimodal contextual information improves future-trajectory prediction for pedestrians, cyclists, scooter riders, and motorcyclists compared with trajectory-history-only models.
Technical contribution
A large experimental comparison of recurrent, attention-based, graph-based, state-space, and probabilistic predictors, with context ablations, temporal analysis, and uncertainty evaluation.
Research highlights
Implemented
06- End-to-end trajectory preprocessing and model-evaluation workflow
- Evaluation across pedestrian, cyclist, scooter, motorcycle, and grouped Cycle+Scooter settings
- Recurrent, Transformer, Mamba/SSM, GCN, and MDN model comparisons
- Trajectory-only and multimodal context-fusion ablations
- Sampling-frequency analysis from 3 Hz to 20 Hz
- ADE, FDE, NLL, predictive-uncertainty, and error-versus-horizon analysis
Research prototypes
03- Agent-interaction modeling variants
- Probabilistic forecasting with Mixture Density Networks
- Contextual-feature and temporal-sampling ablation studies
Next steps
03- Cross-dataset evaluation
- Probabilistic calibration and conformal prediction
- Real-time ADAS inference study
Technical approach
Which neural architecture generalizes best across different VRU groups?
How much do spatial grids, traffic-light state, velocity, and distance contribute?
Can probabilistic predictions and NLL identify forecasts that are likely to have high spatial error?
Research evaluation
Experiments
- Cross-architecture comparison across 7+ model families
- Multimodal ablations: trajectory-only, no-grid, no-traffic-light, no-velocity, and no-distance
- MDN component-count and decoding-strategy studies
- Temporal sampling analysis across 3–20 Hz
- Prediction error across the forecasting horizon
Metrics
- ADE
- FDE
- Negative Log-Likelihood
- Predictive uncertainty
- Pearson correlation
Next research direction
Where this research goes next
Create a fully portable reproduction protocol
Research context
Collaboration