Plane Prior for RGB-D based Visual Odometry

The fusion of low-cost RGB and depth cameras has improved Visual Odometry’s performance significantly especially when encountering feature sparsity and rotational dynamics. However, established Red, Green, Blue, and Depth (RGB-D) methods still lose tracking or accumulate large drift when confronted with strong feature sparsity, as is often the case in indoor pedestrian localization. To address this challenge we propose a novel Plane Prior that extracts the largest planes from successive depth images and computes a probability distribution over the camera’s rotational movement based on the planes’ correspondence. Our method demonstrates resilience even under extremely challenging conditions, such as when the camera is directed towards a uniformly textured surface, by providing crucial motion insights without reliance on texture, multiple planes, or distinct lines. Moreover, it remains unaffected by reflections and shadows, which are common in indoor settings. To demonstrate the effectiveness of our method, we integrate our Plane Prior into the maximum likelihood optimization of the camera pose in the Direct Visual Odometry framework. We show on publicly available as well as self-collected pedestrian data that our Prior significantly reduces positional drift in scenarios of strong feature sparsity.