A deep visual odometry architecture integrating spatial convolution and selective sequence modeling

This paper investigates a monocular visual odometry (VO) method based on deep learning. In recent years, deep learning techniques have demonstrated significant advantages in VO applications, effectively replacing complex engineering steps such as feature extraction and outlier rejection in traditional VO pipelines. To this end, we propose a novel network architecture that integrates ego-motion estimation with sequence-based deep neural network learning. The proposed method employs a Convolutional Neural Network (CNN) to directly extract features from raw RGB image sequences and introduces the Mamba module—based on state space modeling—as a replacement for conventional recurrent neural networks (RNNs) to capture inter-frame dynamics. The network estimates the relative six degrees-of-freedom (6-DoF) camera pose between each frame in the sequence and is capable of implicitly learning the absolute scale without requiring camera intrinsic parameters. This enables full trajectory reconstruction without any post-calibration. Extensive experiments conducted on the KITTI dataset demonstrate that the proposed method achieves superior accuracy and robustness compared to both traditional and other deep learning-based VO approaches.