Development of a Camera Motion Estimation Method Utilizing Motion Blur in Images

Motion blur presents a significant challenge in visual SLAM. Since a satisfying performance highly relies on clear and feature-rich images, the system will easily fail to extract features and lose track due to motion blur. To address this issue, this study introduces a novel solution that estimates the camera motion using blur features extracted from images. The proposed approach models the relationship between camera motion and motion blur, creating a comprehensive motion blur dataset labeled with camera motion. By decoupling the motion estimation process into predicting magnitude and direction separately, a neural network is trained using blur and depth images as inputs to output the camera motion. Experimental results demonstrate that the proposed method successfully estimates motion from blur, even in long-term blur scenarios. This method can potentially serve as an auxiliary motion estimation module to enhance the robustness and accuracy of the visual odometer when motion blur is encountered.