RWKV-VIO: An Efficient and Low-Drift Visual-Inertial Odometry Using an End-to-End Deep Network

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Machine Learning

Background:

• Visual-Inertial Odometry (VIO) is crucial for autonomous navigation.
• Deep learning VIO methods struggle with temporal modeling and computational efficiency.
• Existing models like LSTMs and Transformers are computationally expensive and have limitations in handling temporal scales.

Purpose of the Study:

• To develop a novel VIO framework addressing temporal modeling and computational efficiency challenges.
• Introduce RWKV-VIO, a lightweight and computationally efficient VIO solution.
• Improve feature extraction and utilization of historical inertial data.

Main Methods:

• Proposed RWKV-VIO framework based on the RWKV architecture.
• Implemented a lightweight structure with linear computational complexity.
• Developed a novel Inertial Measurement Unit (IMU) encoder with residual connections and channel alignment.
• Utilized a parallel encoding strategy with two independently initialized encoders for multi-dimensional feature extraction.

Main Results:

• RWKV-VIO demonstrates superior computational efficiency and a lightweight design.
• Significantly reduced model size and inference time compared to advanced methods.
• Achieved top-ranked positioning accuracy on publicly shared datasets.

Conclusions:

• RWKV-VIO effectively addresses the limitations of existing deep learning VIO methods.
• The framework offers a computationally efficient and accurate solution for autonomous navigation.
• The novel IMU encoder and parallel encoding strategy enhance feature extraction and model performance.