A dynamic scene depth estimation method

Abstract

This invention discloses a dynamic scene depth estimation method belonging to the field of autonomous driving perception. Based on binocular event flow, this method first extracts features from the left and right eye images, and then aligns the historical feature space through optical flow estimation and bilinear sampling. Next, it achieves temporal feature fusion based on feature uncertainty, constructs an initial cost volume, and estimates disparity uncertainty. Subsequently, it dynamically adjusts the historical window according to disparity uncertainty, completing cost volume-level adaptive temporal fusion to generate a fused cost volume. Finally, it achieves sub-pixel disparity regression through probability-weighted summation, outputting a continuous disparity map. Training employs a multi-loss collaborative optimization strategy, backpropagating the uncertainty prediction gradient to the feature extraction module to achieve collaborative optimization. By using optical flow alignment and uncertainty-driven fusion, it solves the problems of temporal discontinuity and noise sensitivity in dynamic scene event stereo matching, achieving depth estimation in scenarios with sparse events and motion blur, thus filling the gap in dynamic scene event camera stereo matching technology.

Images