RoNIN: Robust Neural Inertial Navigation in the Wild: Benchmark, Evaluations, and New Methods (1905.12853v1)

Abstract: This paper sets a new foundation for data-driven inertial navigation research, where the task is the estimation of positions and orientations of a moving subject from a sequence of IMU sensor measurements. More concretely, the paper presents 1) a new benchmark containing more than 40 hours of IMU sensor data from 100 human subjects with ground-truth 3D trajectories under natural human motions; 2) novel neural inertial navigation architectures, making significant improvements for challenging motion cases; and 3) qualitative and quantitative evaluations of the competing methods over three inertial navigation benchmarks. We will share the code and data to promote further research.

• The paper introduces a novel dataset with over 42.7 hours of IMU data and robust neural architectures for real-world inertial navigation.
• It refines models including ResNet, LSTM, and TCN with innovative normalization and loss functions to accurately predict trajectories from noisy IMU data.
• The study demonstrates significant reductions in trajectory errors compared to traditional methods, paving the way for reliable mobile and autonomous navigation.

Overview of "RoNIN: Robust Neural Inertial Navigation in the Wild: Benchmark, Evaluations, and New Methods"

The paper presents RoNIN, a set of advancements in the domain of inertial navigation based on data-driven approaches, emphasizing its importance for mobile devices. The authors provide new methodologies along with comprehensive evaluations, setting a substantial benchmark for inertial navigation.

Dataset and Benchmark Establishment

A significant contribution of this paper is the introduction of a novel dataset specifically designed for inertial navigation tasks. The RoNIN dataset includes over 42.7 hours of IMU sensor data from 100 different subjects, vastly expanding upon the scope of prior datasets like OXIOD and RIDI. This dataset showcases a wide range of natural human activities, allowing the development of more robust navigation algorithms that can generalize better to real-world scenarios.

The two-device data collection protocol employed by the authors allows for authentic movement patterns without the constraints usually imposed by older methods. This innovation has permitted the collection of data more representative of typical, everyday device usage scenarios.

Advanced Neural Architectures

RoNIN introduces refined neural network architectures, leveraging models such as ResNet, LSTM, and Temporal Convolutional Networks (TCNs). These architectures are fine-tuned for the task of effectively estimating positional trajectories from IMU data.

1. RoNIN ResNet: Adapting the ResNet-18 framework to process sequential IMU data, this solution focuses on handling large datasets efficiently, balancing performance and computational cost.
2. RoNIN LSTM: This architecture captures long-term dependencies within inertial data, facilitating robust predictions even in complex movement scenarios.
3. RoNIN TCN: Utilizing dilated convolutions, this architecture efficiently handles temporal sequences and exhibits superior performance in sequence prediction tasks over extended periods.

The robustness of these architectures is enhanced by innovative coordinate frame normalization strategies and loss functions, such as latent and strided velocity losses, which significantly improve the neural networks' abilities to handle noisy data and maintain high signal-to-noise ratios.

Evaluation and Results

Evaluations conducted across RIDI, OXIOD, and the newly established RoNIN dataset demonstrate significant performance improvements over existing methods. The paper reports notable reductions in Absolute Trajectory Error (ATE) and Relative Trajectory Error (RTE), with RoNIN consistently outperforming traditional methods such as double integration, PDR, and competing data-driven approaches.

The precision achieved by RoNIN architectures is attributed to their ability to accurately infer velocities and trajectory orientations, a testament to the synergy between the proposed neural models and the comprehensive dataset.

Implications and Future Directions

The findings and methodologies introduced by RoNIN have notable implications for the field of mobile and autonomous navigation. This work paves the way for improved inertial navigation systems capable of functioning reliably under a wide variety of real-world conditions without the need for external aids like GPS.

Future work can explore more robust handling of device orientation dependency, potentially embedding enhanced sensor fusion techniques to address the challenges posed by noisy or incorrect orientation data. Researchers are also encouraged to extend these findings to more challenging environments and test scenarios, utilizing the openly shared datasets and code to foster innovation in ubiquitous navigation systems.

The contributions of this paper mark an important step towards achieving universally reliable and autonomous navigation solutions in consumer devices, with practical applications spanning augmented reality, autonomous systems, and human-centric computing interfaces. Through the RoNIN dataset and its associated methodologies, the paper establishes a robust platform for further research in the field of inertial navigation.