BatchEnsemble-based perception uncertainty quantification for autonomous vehicles

Abstract

Quantifying uncertainty significantly enhances the reliability of perception in autonomous vehicles and provides more comprehensive environmental information for downstream modules. However, most existing perception methods lack the capacity to effectively estimate the associated uncertainty. To address this gap, we propose a BatchEnsemble-based network for uncertainty quantification in 3D object detection using point cloud data. Specifically, a BatchEnsemble-based convolutional layer is designed to reduce the memory overhead associated with ensemble-based paradigms. Building upon this, a series of probabilistic object detection networks are constructed by directly modeling object attributes using multivariate Gaussian distributions, thereby enabling the parallel extraction of both object features and their associated variances. Subsequently, an uncertainty-aware fusion strategy is introduced to integrate and filter multiple detection results based on an uncertainty quantification metric—namely, the Uncertainty Index—thereby yielding more reliable and comprehensive outputs. The proposed method is validated on the KITTI dataset. Experimental results demonstrate its competitive accuracy performance and effectiveness across various scenarios, including objects of differing detection difficulty, identification of false-positive results, and under adverse conditions such as snowy weather and sensor degradation.