Uncertainty propagation from sensor data to deep learning models in autonomous driving

Context: Deep learning has been widely used in Autonomous Driving Systems (ADS). Though significant progress has been made regarding their efficiency and accuracy, uncertainty remains a critical factor affecting ADS safety. Such uncertainties are often due to environmental noise and/or imperfect algorithm structures. Studies on uncertainty quantification mostly focus on single classification tasks and overlook how uncertainties propagate from the perception to downstream decision-making, studying of which is critical, as the interplay between perception and decision-making can significantly impact the overall safety of ADS. Objectives: We quantify and understand the uncertainty propagation from sensor data to deep learning models, as well as its impact on ADS safety. Methods: We present an empirical study that quantifies both aleatoric and epistemic uncertainties and assesses how such uncertainties propagate and impact ADS safety under various sensor noise conditions. We also investigate the suitability of two epistemic uncertainty quantification methods (i.e., MC Dropout and Deep Ensembles) to ADS tasks and their cost-effectiveness in selecting highly-uncertain samples. Results: Results show that increased noise can significantly increase uncertainty and degrade model performance, thereby compromising decision-making and potentially impacting ADS safety. Both MC Dropout and Deep Ensembles effectively measure the model's epistemic uncertainty, with MC Dropout showing higher correlation with ADS safety, and saving time and computational costs. Moreover, there are significant differences in the highly-uncertain samples they identified. Conclusion: Our results show the importance of considering uncertainty propagation to ensure the ADS safety. Compared to Deep Ensembles, MC Dropout's efficiency makes it a more suitable choice in the context of ADS.