Efficient Collaborative Perception With Integrated Uncertainty Estimation via Evidence Regression

The collaborative exchange of information between multiple agents effectively addresses the limitations of individual perception. However, in practical applications, agents often lack the ability to estimate uncertainty for detected targets, which can negatively impact decision-making and control processes. As a result, uncertainty estimation is crucial for multi-agent systems. In this paper, we introduce Evidence Regression for Collaborative Perception (ER-CoPe), a novel and efficient collaborative perception framework that employs evidential deep learning to explicitly quantify and manage aleatoric and epistemic uncertainties. ER-CoPe incorporates an Evidence Regression (ER) module that models bounding box parameter uncertainties using a Normal-Inverse Gamma distribution, effectively eliminating the computational overhead associated with traditional repeated inference-based uncertainty quantification. Additionally, we propose gradient regularization to mitigate the gradient vanishing issue in regions of high uncertainty and introduce uncertainty regularization to address evidence shrinkage. These novel regularization techniques are integrated into a multi-level uncertainty loss function we designed to maintain training stability. Furthermore, we propose a multi-task loss function that dynamically balances detection accuracy and uncertainty estimation during the training process. Extensive experiments conducted on two widely-used collaborative perception datasets, OPV2V and V2X-Set, validate that ER-CoPe achieves state-of-the-art performance, significantly improving both the reliability and efficiency of multi-agent collaborative perception systems, making it particularly suitable for autonomous driving applications.