A Normal Distribution Transform-Based Radar Odometry Designed For Scanning and Automotive Radars (2103.07908v3)

Abstract: Existing radar sensors can be classified into automotive and scanning radars. While most radar odometry (RO) methods are only designed for a specific type of radar, our RO method adapts to both scanning and automotive radars. Our RO is simple yet effective, where the pipeline consists of thresholding, probabilistic submap building, and an NDT-based radar scan matching. The proposed RO has been tested on two public radar datasets: the Oxford Radar RobotCar dataset and the nuScenes dataset, which provide scanning and automotive radar data respectively. The results show that our approach surpasses state-of-the-art RO using either automotive or scanning radar by reducing translational error by 51% and 30%, respectively, and rotational error by 17% and 29%, respectively. Besides, we show that our RO achieves centimeter-level accuracy as lidar odometry, and automotive and scanning RO have similar accuracy.

• The paper presents an NDT-based radar odometry framework that integrates thresholding, probabilistic submap building, and NDT scan matching.
• It achieves a 51% reduction in automotive translational error and up to a 30% reduction in scanning radar errors, enhancing real-time navigation accuracy.
• The approach is computationally efficient and robust to noise, offering potential for sensor fusion in autonomous navigation systems.

A Normal Distribution Transform-Based Radar Odometry for Scanning and Automotive Radars

This paper presents a versatile radar odometry methodology designed to operate seamlessly with both automotive and scanning radars, a notable innovation in the field of robotics and autonomous navigation systems. By leveraging the Normal Distribution Transform (NDT) for scan matching, the authors propose an efficient radar odometry (RO) framework that is equipped to handle the noise-dense environment typical of radar images while maintaining robustness to outliers.

Methodological Contributions

The radar odometry approach comprises three primary components: thresholding, probabilistic submap building, and NDT-based scan matching. Each plays a critical role in processing the discrepancies of data arising from different types of radars.

1. Thresholding: This pre-processing step effectively removes noise from radar images, a major challenge when working with scanning radars, which provide rich but often noisy data. The authors use a fixed threshold, alleviating the need for complex noise masking networks, to ensure only data with significant signal return is considered.
2. Probabilistic Submap Building: The submap is a sophisticated data fusion technique that aggregates multiple radar scans while incorporating measurement uncertainties, thereby enhancing data density especially for the sparse outputs from automotive radars. Importantly, this submap accounts for the ego-motion, integrating velocity estimates to construct a coherent environmental map over time.
3. NDT-Based Scan Matching: At the crux of the approach is the NDT, which converts radar point clouds into probability distributions. This transformation allows for precise alignment of consecutive radar scans, reducing both translational and rotational errors. The use of weighted P2D NDT further refines the matching process by emphasizing higher-confidence measurements.

Empirical Evaluation and Results

The empirical validation was conducted using two comprehensive datasets: the Oxford Radar RobotCar for scanning radar data and the nuScenes dataset for automotive radar. Benchmarking against state-of-the-art methods, the proposed RO demonstrated significant improvements in both scenarios:

• For automotive radars, the method achieved a 51% reduction in translational error and a 17% reduction in rotational error compared to the previous best approaches.
• For scanning radars, results were similarly impressive, achieving a 30% reduction in translational error and a 29% reduction in rotational error.

The effectiveness of the proposed approach is further underscored by its performance relative to lidar-based odometry, achieving centimeter-level accuracy, despite inherent differences in sensor modalities.

Implications and Future Directions

The paper's outcomes suggest broad implications for the development of robust autonomous navigation systems. By providing a unified framework applicable to both radar types, this work pushes forward the potential for radar-based navigation in diverse environmental conditions where traditional optical sensors may fail. The proposed framework does not rely on complex training or deep learning models, favoring instead an elegant and computationally efficient solution conducive to real-time applications.

Looking ahead, future research may explore the fusion of these RO methods with additional sensor modalities such as IMUs to further enhance accuracy and robustness. There's also potential for extending the methodology to 4D radar systems to support 6-DoF odometry, which could considerably improve the functionality of autonomous vehicles and robots in complex, real-world environments.