Uncertainty Quantification of Collaborative Detection for Self-Driving

Submitted to ICRA 2023

1Department of Computer Science and Engineering, University of Connecticut
2Tandon School of Engineering, New York University

Video




Left figure shows detection results of intermediate collaboration in bird's eye view (BEV), and right figure zooms on a specific part to show the robust range of two detections. Red boxes are predictions, and green boxes are ground truth. The orange ellipse denotes the covariance of each corner. The shadow convex hull shows the uncertainty set of the detected object. The shadow convex hull covers the green bounding box in most cases, which helps the later modules in self-driving tasks, such as trajectory prediction with uncertainty propogation and robust planning and control. With our Double-M Quantification method, detected objects with low accuracy tend to have large uncertainties.

Abstract

Sharing information between connected and autonomous vehicles (CAVs) fundamentally improves the performance of collaborative object detection for self-driving. However, CAVs still have uncertainties on object detection due to practical challenges, which will affect the later modules in self-driving such as planning and control. Hence, uncertainty quantification is crucial for safety-critical systems such as CAVs. Our work is the first to estimate the uncertainty of collaborative object detection. We propose a novel uncertainty quantification method, called Double-M Quantification, which tailors a moving block bootstrap (MBB) algorithm with direct modeling of the multivariant Gaussian distribution of each corner of the bounding box. Our method captures both the epistemic uncertainty and aleatoric uncertainty with one inference based on the offline Double-M training process. And it can be used with different collaborative object detectors. Through experiments on the comprehensive CAVs collaborative perception dataset, we show that our Double-M method achieves up to 4.09 times improvement on uncertainty score and up to 3.13% accuracy improvement, compared with the state-of-the-art uncertainty quantification. The results also validate that sharing information between CAVs is beneficial for the system in both improving accuracy and reducing uncertainty.

Contribution

  • To the best of our knowledge, our proposed Double-M Quantification is the first attempt to estimate the uncertainty of collaborative object detection. Our method tailors a moving block bootstrap algorithm to estimate both the epistemic and aleatoric uncertainty in one inference.
  • We design a novel loss function and the representation format of the bounding box uncertainty in the direct modeling component to estimate the aleatoric uncertainty.
  • We validate the advantages of the proposed methodology based on V2X-SIM [1] and show that our Double-M Quantification method reduces the uncertainty and improves the accuracy.
  • We also show that sharing information between CAVs is beneficial for the system in both improving accuracy and reducing uncertainty.

  • Method



    Implementation of our Double-M Quantification method on collaborative object detection. (a) Early collaboration shares raw point cloud with other agents, and (b) intermediate collaboration shares intermediate feature representations with other agents. (c) Double-M Quantification method estimates the multi-variate Gaussian distribution of each corner. Our Double-M Quantification method can be used on different collaborative object detection. During the training stage, Double-M Quantification tailors a moving block bootstrapping algorithm to get the final model parameter, the average aleatoric uncertainty of the validation dataset and the covariance of all residual vectors for epistemic uncertainty. During the inference stage, we can combine the aforementioned two covariance matrix and the predicted covariance matrix from the object detector to compute the final covariance matrix.


    Qualitative  Results

     


    Visualization of our Double-M Quantification results on different scenes of V2X-Sim [1]. The results of LB, DN, and UB are respectively shown in the first, second, and third row. Red boxes are predictions, and green boxes are ground truth. Orange ellipse denotes the covariance of each corner. We can see our Double-M Quantification predicts large orange ellipses when the differece between the red bounding box and the corresponding green bounding box is huge, which means our method is efficient.

    BibTeX

    @article{Su2022uncertainty,
          author    = {Su, Sanbao and Li, Yiming and He, Sihong and Han, Songyang and Feng, Chen and Ding, Caiwen and Miao, Fei},
          title     = {Uncertainty Quantification of Collaborative Detection for Self-Driving},
          year={2022},
          eprint={2209.08162},
          archivePrefix={arXiv}
      }