Deep Extended Measurement Models for ADAS
Master's thesis on learning the measurement model for vehicle tracking from data instead of hand-crafting it.
Background
Autonomous vehicles track the vehicles around them using a Bayesian filter, and a key part of that filter is the measurement model: how sensor detections relate to an object's state. Hand-crafted models need expert knowledge, lots of tuning and simplifying assumptions. With large autonomous-driving datasets now available, the model can be learned from data instead.
Approach
- Modelled measurements as a Poisson Point Process, learning its intensity function and rate.
- Compared two model families: parametric Gaussian Mixture Models and Energy-Based Models.
- Trained and evaluated on NuScenes trajectory data, with metrics covering a range of driving scenarios.
- Integrated the best model into a multi-object tracking particle filter and benchmarked it against the Variational Radar Model (VRM).
Findings
- The energy-based models performed on par with or better than the established VRM in the tested particle filter, which makes them a viable way to represent measurement models in autonomous systems.
- EBMs are more flexible and use less condensed object-state information than the parametric approaches.
- Further validation with other filters and broader datasets is the natural next step.