<p>Visual odometry (VO) is a critical component in visual simultaneous localization and mapping systems, enabling the estimation of camera motion from sequences of images. However, these deep learning-based solutions often face challenges related to insufficient generalization across different environments and high computational resource requirements. To address these issues, this paper proposes an end-to-end, lightweight deep learning framework for VO. It integrates a pose regression network with an optical flow estimation network, aiming to achieve efficient processing while maintaining accuracy. By incorporating depth information, our method effectively mitigates the scale ambiguity issue commonly encountered in monocular VO systems, thereby enhancing the accuracy and reliability of the estimated camera motion. Specifically, by leveraging depth data obtained from a monocular depth estimation network, our method demonstrates improved consistency in scale estimation and enhanced robustness across varying environments. Moreover, the reduced computational overhead allows for realtime operation on platforms with limited resources.</p>