Deep Learning Approaches in Monocular Visual Odometry

Monocular visual odometry (VO) seeks to estimate a camera’s trajectory using only sequential images from a single lens. Traditional methods rely on hand-crafted features and geometric optimisation, but are prone to scale drift, failure in low-texture scenes and challenges in dynamic environments. Deep learning approaches have transformed the field by learning image representations and motion models directly from data. Convolutional neural networks extract hierarchical features for frame-to-frame pose estimation, while recurrent modules capture temporal consistency. Unsupervised frameworks jointly infer depth and motion without ground-truth labels, exploiting photometric and geometric consistency. More recent architectures integrate transformer-style self-attention to model long-range dependencies and global context, thereby reducing drift over extended trajectories. Parallel work on multi-scale modelling and hybrid networks addresses the inherent scale ambiguity of monocular systems, combining learned priors with classical constraints to improve robustness. These advances have enabled real-time, end-to-end pipelines that rival geometry-based state-of-the-art methods, with broad implications for autonomous vehicles, robotics and augmented reality around the globe.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent contributions have demonstrated the power of attention mechanisms and unsupervised losses in monocular VO. A transformer-based fusion network replaces standard convolutional encoders with space-time self-attention and cross-frame interaction, yielding marked reductions in trajectory drift on benchmark datasets. This model learns relative poses between adjacent frames while a global subnetwork refines long-range dependencies, outperforming purely convolutional baselines in both translational and rotational accuracy. Unsupervised multi-scale frameworks have also emerged, leveraging densely linked atrous convolutions and non-local attention to capture features at diverse spatial scales. By integrating multi-scale information, these systems achieve state-of-the-art rotation estimation and depth prediction without requiring ground-truth supervision. Complementary work on rotational adjustment networks employs a model-free epipolar constraint to correct the rotation predicted by a learnt pose network. By matching deep keypoints and applying an iterative frame-to-frame solver during training and inference, such hybrid designs deliver enhanced rotational accuracy and generalise robustly across motion patterns. Together, these studies illustrate a trend towards blending learned priors, self-supervision and geometric reasoning to address the core challenges of monocular VO.

Deep Learning Approaches in Monocular Visual Odometry publication trend

The graph below shows the total number of articles in deep learning approaches in monocular visual odometry across all publications each year (not limited to Nature Index journals).

Technical terms

Monocular visual odometry: Estimation of camera motion and trajectory from a sequence of images captured by a single camera.

Convolutional neural network: A deep learning architecture that applies spatially localised filters to extract hierarchical image features.

Transformer: A neural architecture relying on self-attention mechanisms to model global dependencies within inputs.

Unsupervised learning: A training paradigm where models infer structures or tasks from data without explicit ground-truth labels.

Scale drift: The gradual accumulation of scale errors in monocular VO due to the inability to recover absolute distance from a single view.

Epipolar geometry: The geometric relationship between two camera views used to constrain correspondences and recover relative motion.

References

1. A Global Pose and Relative Pose Fusion Network for Monocular Visual Odometry. IEEE Access (2024).
2. An Unsupervised Monocular Visual Odometry Based on Multi-Scale Modeling. Sensors (2022).
3. RAUM-VO: Rotational Adjusted Unsupervised Monocular Visual Odometry. Sensors (2022).