BEV-DWPVO: BEV-based Differentiable Weighted Procrustes for Low Scale-drift Monocular Visual Odometry on Ground

TL;DR

This work tackles scale drift in monocular visual odometry by grounding pose estimation in a unified Bird's-Eye View (BEV) representation, which enables reducing the pose from $6$-DoF to $3$-DoF under a ground-plane assumption. It introduces BEV-DWPVO, an end-to-end differentiable pipeline that maps perspective images to BEV via a PV-BEV encoder, extracts and matches BEV keypoints, and solves the relative pose with a differentiable Weighted Procrustes solver, trained with pose supervision only. Across NCLT, Oxford, and KITTI, BEV-DWPVO achieves superior scale consistency and pose accuracy compared to perspective-view methods and other BEV baselines, while maintaining real-time performance. The work demonstrates that BEV representations coupled with geometric constraints provide robust, interpretable, and scalable monocular VO for ground vehicles, with potential for extension to non-planar terrains and multi-layer BEV representations.

Abstract

Monocular Visual Odometry (MVO) provides a cost-effective, real-time positioning solution for autonomous vehicles. However, MVO systems face the common issue of lacking inherent scale information from monocular cameras. Traditional methods have good interpretability but can only obtain relative scale and suffer from severe scale drift in long-distance tasks. Learning-based methods under perspective view leverage large amounts of training data to acquire prior knowledge and estimate absolute scale by predicting depth values. However, their generalization ability is limited due to the need to accurately estimate the depth of each point. In contrast, we propose a novel MVO system called BEV-DWPVO. Our approach leverages the common assumption of a ground plane, using Bird's-Eye View (BEV) feature maps to represent the environment in a grid-based structure with a unified scale. This enables us to reduce the complexity of pose estimation from 6 Degrees of Freedom (DoF) to 3-DoF. Keypoints are extracted and matched within the BEV space, followed by pose estimation through a differentiable weighted Procrustes solver. The entire system is fully differentiable, supporting end-to-end training with only pose supervision and no auxiliary tasks. We validate BEV-DWPVO on the challenging long-sequence datasets NCLT, Oxford, and KITTI, achieving superior results over existing MVO methods on most evaluation metrics.

Figures (4)

• Figure 1: ATE metrics for different methods on NCLT and Oxford datasets. We compare the ratio of ATE metrics for different methods under SE(3) alignment and Sim(3) alignment using the composite metric $\log_2(\text{SE(3)}/\text{Sim(3)})$. This comparison demonstrates the scale drift of different methods. Our method uses BEV representation and differentiable weighted Procrustes, achieving the best scale consistency and the lowest average ATE metrics across both datasets.
• Figure 2: Overview of the proposed framework.
• Figure 3: Trajectory and keypoint matching visualization. The NCLT dataset uses a forward monocular camera, while the Oxford dataset uses a rear monocular camera. Keypoint-matching images are cropped to show relevant areas within the field of view.
• Figure 4: Logarithmic scale factor variation along the path.