Abstract
Accurate and robust pose estimation plays a crucial role in many robotic systems. Popular algorithms for pose estimation typically rely on high-fidelity and high-frequency signals from various sensors. Inclusion of these sensors makes the system less affordable and much more complicated. In this work we introduce a novel approach for the robotic odometry which only requires a single camera and, importantly, can produce reliable estimates given even extremely low-frequency signal of around one frame per second. The approach is based on matching image features between the consecutive frames of the video stream using deep feature matching models. The resulting coarse estimate is then adjusted by a convolutional neural network, which is also responsible for estimating the scale of the transition, otherwise irretrievable using only the feature matching information. We evaluate the performance of the approach in the AISG-SLA Visual Localisation Challenge and find that while being computationally efficient and easy to implement our method shows competitive results with only around $3^{\circ}$ of orientation estimation error and $2m$ of translation estimation error taking the third place in the challenge.
Introduction
Most visual odometry (VO) and SLAM algorithms are designed for high-frequency data streams, typically 30 fps or higher. These systems often rely on auxiliary sensors like IMUs or LiDAR to resolve scale ambiguity and maintain tracking during rapid motion. However, many real-world robotic applications operate under constraints such as low-frequency data (around 1 fps), unstable connections, and limited hardware (single RGB camera). In such scenarios, traditional sparse feature tracking often fails due to the large displacement between frames. This paper introduces "Match and Locate," a two-stage framework designed to provide robust monocular odometry at extremely low frequencies.
Methodology
The proposed pipeline consists of two primary stages: a geometric coarse matching stage and a deep learning-based refinement stage.
Stage 1: Coarse Matching-based Localization
To handle the large viewpoint shifts associated with low-frequency capture, the authors utilize LoFTR (Local Feature TRansformer). Unlike traditional sparse detectors like ORB or SIFT, LoFTR performs dense matching, which is significantly more robust when image overlap is minimal.
Following matching, RANSAC is employed to estimate the relative rotation and a unit translation vector. In cases where matching fails due to extreme conditions, the system defaults to a "constant rotation" heuristic, assuming a u-turn based on the specific constraints of the navigation environment.
Stage 2: CNN Refinement
The coarse geometric estimate is then passed to a Convolutional Neural Network (CNN) for refinement. This stage is critical for two reasons: it corrects errors in the initial estimate and, most importantly, it predicts the absolute scale of the translation in meters—a value that cannot be derived from monocular feature matching alone.
To enhance the CNN's performance, the input is augmented with auxiliary data:
- Monocular Depth Maps: Generated via MiDaS.
- Optical Flow: Estimated using RAFT.
- Temporal Data: Time interval encodings between frames.
Experimental Results
The framework was evaluated using the AISG–SLA Visual Localisation Challenge dataset. The "Match and Locate" method achieved 3rd place, demonstrating its effectiveness in real-world, low-frequency scenarios.
Performance and Ablation
The system achieved an orientation error of approximately 3° and a translation error of 2.0m. An ablation study highlights the necessity of the two-stage approach. As shown in the table below, a model using only the second stage (CNN-only) suffers from high translation errors (11.0m–12.0m). Integrating the geometric hints from Stage 1 reduces this error to 2.0m, representing a 5x improvement in accuracy.
Computational Efficiency
Despite the complexity of the deep learning models, the system processes a frame pair in approximately 1 second on an NVIDIA RTX 3090. This makes it suitable for near-real-time operation at the target frequency of 1 fps.
Conclusion
The "Match and Locate" framework demonstrates that robust monocular odometry is possible even at very low frame rates by combining dense feature matching with CNN-based refinement. By leveraging auxiliary geometric maps (depth and flow), the model successfully overcomes scale ambiguity. Future work will focus on improving computational efficiency for lower-end hardware and implementing more sophisticated temporal modeling to handle extended signal losses.
Summary updated