DOU-Pose: Robust Camera-Based Visual Localization for Autonomous Vehicles in Repetitive and Low-Texture Intelligent Transportation Environments

Abstract

Accurate and robust vehicle localization is essential for autonomous driving. However, existing visual pose estimation methods often struggle in scenarios dominated by repetitive structures or sparse textures. These conditions lead to ambiguous predictions of 3D scene coordinates and a high proportion of structured outliers—erroneous predictions forming coherent clusters that deceive standard estimators. To address these limitations, this paper proposes DOU-Pose (Depthwise Over-parameterized U-shaped Pose estimation), a visual pose estimation framework built upon the Differentiable SAmple Consensus (DSAC)* pipeline. The core idea is to enhance the discriminative capability of scene coordinate regression through improved feature extraction. Specifically, we replace standard convolutional layers with Depthwise Over-parameterized Convolution (DO-Conv), which introduces auxiliary learnable depthwise kernels during training to enrich the representational capacity of the network, while allowing their fusion into a single kernel for inference. Furthermore, a U-shaped regression network with transposed convolutions is designed to preserve spatial details and strengthen fine-grained geometric reasoning. The entire pipeline is trained end-to-end by coupling dense scene coordinate prediction with a differentiable robust estimator. Extensive experiments demonstrate that DOU-Pose achieves competitive performance on public benchmarks and clear robustness improvements on the self-collected Campus-AV dataset, especially in repetitive and low-texture outdoor driving scenarios.