AKG‐VO: Adaptive Keyframe Generation Method for Improving Visual Odometry in Autonomous Vehicles

Abstract

Keyframe selection remains a critical challenge in visual odometry, significantly influencing overall system performance. Traditional approaches have primarily focused on identifying meaningful frames; however, in cases where the robot operates at high velocities or sensor data is temporarily unavailable, large interframe gaps can result in an insufficient number of keyframes. To address this limitation, a novel methodology is proposed that not only refines keyframe selection but also adaptively generates additional keyframes through video frame interpolation. The proposed approach estimates a sufficient number of images, enabling the adaptive generation of keyframes. Additionally, a loss function is introduced to minimize motion blur, facilitating the generation of keyframes more suited for visual odometry. Experimental evaluations demonstrate the effectiveness of the method, yielding superior performance across various trajectory evaluations, thereby validating the robustness and accuracy of the proposed approach in enhancing visual odometry systems.