Defensive Perception: Estimation and Monitoring of Neural Network Performance under Deployment

TL;DR

This paper introduces a method for estimating and monitoring neural network performance in autonomous driving, using uncertainty envelopes to detect domain shifts and ensure safety without modifying the deployed model.

Contribution

It presents a novel uncertainty estimation envelope based on Monte Carlo Dropout that monitors neural network performance during deployment without requiring model modifications.

Findings

Effective detection of domain shifts such as night, rain, snow.

Reduces computational costs and estimation noise.

Enables safety triggers and performance notifications during deployment.

Abstract

In this paper, we propose a method for addressing the issue of unnoticed catastrophic deployment and domain shift in neural networks for semantic segmentation in autonomous driving. Our approach is based on the idea that deep learning-based perception for autonomous driving is uncertain and best represented as a probability distribution. As autonomous vehicles' safety is paramount, it is crucial for perception systems to recognize when the vehicle is leaving its operational design domain, anticipate hazardous uncertainty, and reduce the performance of the perception system. To address this, we propose to encapsulate the neural network under deployment within an uncertainty estimation envelope that is based on the epistemic uncertainty estimation through the Monte Carlo Dropout approach. This approach does not require modification of the deployed neural network and guarantees expected model performance. Our defensive perception envelope has the capability to estimate a neural network's performance, enabling monitoring and notification of entering domains of reduced neural network performance under deployment. Furthermore, our envelope is extended by novel methods to improve the application in deployment settings, including reducing compute expenses and confining estimation noise. Finally, we demonstrate the applicability of our method for multiple different potential deployment shifts relevant to autonomous driving, such as transitions into the night, rainy, or snowy domain. Overall, our approach shows great potential for application in deployment settings and enables operational design domain recognition via uncertainty, which allows for defensive perception, safe state triggers, warning notifications, and feedback for testing or development and adaptation of the perception stack.