Reliable Autonomous Driving Environment Perception: Uncertainty Quantification of Semantic Segmentation

Abstract

Despite the impressive achievements of computer vision technologies such as semantic segmentation, their applications in safety-critical areas, such as autonomous driving, present substantial challenges, particularly in ensuring the safety of the intended functionality (SOTIF). It is well-recognized that the lack of confidence estimation or overconfidence in a model prediction hinders model applicability and dependability in critical sectors. Profiting from the expressive modeling ability of Dempster–Shafer theory for uncertain information, we propose EviSeg as an approach of uncertainty estimation for semantic segmentation models, which is grounded in the evidential classifier framework. Specifically, we first transform the fully convolutional neural networks used for semantic segmentation via pixelwise classification into an evidential model. Subsequently, the outputs of the penultimate convolutional layer and parameters of the final convolutional layer of a conventionally trained semantic segmentation model constitute a raw evidence pool. Reasoning from this evidence pool, we quantify the predictive uncertainties with the metric conflict. The proposed method does not affect model performance because it does not necessitate alterations to the model architecture and training objective. We utilize the CamVid urban road scene data set and Nighttime Driving data set for our experimental analysis. These experiments demonstrated that, in comparison with the baseline methods, our proposed approach not only provides competitive performance but also enhances computational efficiency significantly. Our study directly contributes to improving the safety and reliability of connected and automated vehicles (CAVs). Such a contribution is crucial to reduce the accidents of CAVs caused by environment perception issues and improving the SOTIF of CAVs.