Event-Based Stereo Depth Estimation With Motion Guidance and Left–Right Consistency

摘要

Depth estimation, which directly captures the structure of observable environmental surfaces, plays a critical role in vision-based applications such as measurement, mapping, autonomous driving, and robot navigation. Specifically, event camera-based stereo depth estimation provides a novel solution to challenging conditions like rapid motion and extreme illumination variations. Currently, deep learning has become the dominant approach for event camera-based stereo depth estimation. However, these methods fail to fully exploit temporal cues in the event stream, resulting in suboptimal event representation clarity. Furthermore, there is room for further reduction in pixel shifts in feature maps before constructing the cost volume. This paper presents a novel event-based stereo depth estimation method with motion guidance and left-right consistency. First, an edge-aware aggregation module (EAA) is proposed, which integrates event frames with motion confidence maps to generate a novel high-definition event representation. Then, a motion-guided attention (MGA) module is introduced, which leverages deformable transformer encoders guided by motion confidence maps to refine edge precision in feature maps. Finally, a census left-right consistency loss function is designed to enhance the left-right consistency of stereo event representation. Experiments conducted in challenging real-world driving scenarios demonstrate that the proposed method outperforms state-of-the-art methods in terms of mean absolute error (MAE) and root mean square error (RMSE) metrics.