Abstract: To address the limitations of traditional visual odometry in terms of accuracy and robustness in complex, dynamic, and low-texture environments, this paper proposes an optimization method based on the variational approach. The method enhances the texture feature representation of RGB images through variational modeling, thereby improving the robustness of feature extraction and matching, while also optimizing the computational efficiency of the local bundle adjustment (BA) in the backend. First, a variational model is constructed to compensate for scattering effects on RGB images, achieving backscatter suppression, adaptive separation of high- and low-frequency noise, and forward-scatter component decomposition, which effectively improves the accuracy of image feature extraction. Second, to address the issues of low contrast and color deviation caused by uneven illumination, a histogram equalization method is introduced for adaptive correction, combined with a frequency-domain decomposition strategy to avoid damaging the continuity of image pixels. Furthermore, a dynamic local BA optimization strategy based on variational inference is adopted, and an adaptive keyframe selection mechanism is employed to enhance optimization efficiency, particularly by incorporating environmental information metrics to ensure improved computational performance. To validate the effectiveness of the proposed method, systematic experimental evaluations are conducted on several public datasets, including TUM, EuRoC, and KITTI. Experimental results demonstrate that the proposed method significantly outperforms existing traditional approaches in terms of both trajectory estimation accuracy and computational efficiency, confirming its advantages in complex environments.