Uncertainty-Aware AB3DMOT by Variational 3D Object Detection

TL;DR

The paper addresses robust 3D object detection and tracking under uncertainty for autonomous driving. It introduces a Variational TANet detector that outputs mean predictions and uncertainty, and integrates this with an uncertainty-aware AB3DMOT via a linear transformation of the predicted uncertainty $\hat{\Sigma} = \alpha I + \beta \Sigma$ to serve as Kalman filter measurement noise. It presents two uncertainty pathways—internal variance and external covariance—and shows external covariance yields stronger tracking gains, with transfer learning from a pretrained TANet to IVTANet providing the best performance. Experiments on KITTI track demonstrate improved MOTA and F1 scores, validating the practical impact of uncertainty-aware 3D MOT and the value of transferring existing models to variational variants.

Abstract

Autonomous driving needs to rely on high-quality 3D object detection to ensure safe navigation in the world. Uncertainty estimation is an effective tool to provide statistically accurate predictions, while the associated detection uncertainty can be used to implement a more safe navigation protocol or include the user in the loop. In this paper, we propose a Variational Neural Network-based TANet 3D object detector to generate 3D object detections with uncertainty and introduce these detections to an uncertainty-aware AB3DMOT tracker. This is done by applying a linear transformation to the estimated uncertainty matrix, which is subsequently used as a measurement noise for the adopted Kalman filter. We implement two ways to estimate output uncertainty, i.e., internally, by computing the variance of the CNN outputs and then propagating the uncertainty through the post-processing, and externally, by associating the final predictions of different samples and computing the covariance of each predicted box. In experiments, we show that the external uncertainty estimation leads to better results, outperforming both internal uncertainty estimation and classical tracking approaches. Furthermore, we propose a method to initialize the Variational 3D object detector with a pretrained TANet model, which leads to the best performing models.

Figures (2)

• Figure 1: Examples of 3D bounding boxes with raw uncertainties obtained by the proposed Variational TANet (top figure), and the uncertainties provided to the Kalman filter of the proposed Uncertainty-aware AB3DMOT (bottom figure). Yellow color represents higher probability, while the purple color represents lower probability. Blue boxes correspond to the mean predictions.
• Figure 2: The pipeline of the proposed Uncertainty-Aware AB3DMOT by Variational 3D Object Detection method. Following TANet liu2019tanet, the input point cloud is voxelized to create a pillar map and processed by a Stacked Triple Attention Module to generate a 2D pseudo-image. This pseudo image is used by a Variational version of the Coarse-to-Fine convolutional network to create pillar-wise predictions. Uncertainty from these predictions is estimated by combining outputs of multiple applications of this network to the same inputs in one of two ways. The Internal Variance approach computes variance of pillar-wise predictions and propagates it through the post-processing steps, while the External Covariance approach computes the final prediction for each sample and then applies an association algorithm to find groups of object detections corresponding to the same object to compute the covariance between them. Uncertain object detections are passed to an Uncertainty-Aware AB3DMOT, which applies a linear transformation to the uncertainty and uses it as an additional value to a Kalman filter. A Kalman filter is used to predict changes in object positions between frames, and Hungarian algorithm is used to associate new object detections with known tracklets.