AR-CP: Uncertainty-Aware Perception in Adverse Conditions with Conformal Prediction and Augmented Reality For Assisted Driving

Doula, Achref; Mühlhäuser, Max; Sanchez Guinea, Alejandro

Open Access

AR-CP: Uncertainty-Aware Perception in Adverse Conditions with Conformal Prediction and Augmented Reality For Assisted Driving

Files

AR_CP_PlotsGeneration.ipynb (data processing and plot generation code) (75.62 KB)

Example from study videos.zip (example videos from study) (43.94 MB)

CVPRW__AR_CP.pdf (Paper preprint) (2.63 MB)

Date

2024-04-30

Type

Software

Authors

Doula, Achref
Mühlhäuser, Max
Sanchez Guinea, Alejandro

Description

Deep learning models play a crucial role in improving driver assistance systems and environmental perception. However, their tendency toward overconfident predictions poses risks to driver safety, particularly in adverse conditions. To address this, we propose AR-CP, an uncertainty-aware framework integrating conformal prediction and augmented reality (AR). AR-CP starts with a conformal prediction step, generating an uncertainty-aware prediction set. Then, AR simplifies and clarifies the visualization of the closest common parent class, reducing misinformation. We present a rigorous formulation and theoretical analysis, evaluating AR-CP on the ROAD dataset. Results demonstrate superior performance compared to existing methods, ensuring safer driving experiences with reduced mental load and heightened situation awareness, as validated by an immersive user study involving 15 participants.