Towards Perception for Autonomy with 4D mmWave Radar : Learning Ego-Motion, Place Recognition, and Uncertainty-Aware Sensor Fusion
Rai, Prashant Kumar (2026)
Tampere University
2026
Teknisten tieteiden tohtoriohjelma - Doctoral Programme in Engineering Sciences
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Väitöspäivä
2026-02-13
Julkaisun pysyvä osoite on
https://urn.fi/URN:ISBN:978-952-03-4390-3
https://urn.fi/URN:ISBN:978-952-03-4390-3
Tiivistelmä
Autonomous systems operate in a range of environments where traditional perception sensors, such as optical cameras and light detection and ranging units, often fail due to adverse visibility or signal loss. These limitations are common in logistics, mining, and construction sectors, where machines function in degraded visual and satellite-denied conditions. This dissertation investigates millimeter-wave imaging radar as a primary sensor for autonomous perception, focusing on three core tasks: ego-motion estimation, place recognition, and radar-inertial fusion.
The proposed methods rely on high-resolution four-dimensional radar signal data and eliminate the need for handcrafted feature extraction or point cloud generation. Instead, neural networks are used to learn motion and spatial features directly from raw radar measurements. First, ego-motion is estimated from three-dimensional heatmap radar data using a convolutional network with attention mechanisms. Second, a representation learning framework is presented for place recognition using raw heatmaps, trained in a self-supervised manner through ego-motion as a proxy task. Third, an uncertainty-aware fusion approach integrates radar-based velocity estimates with inertial measurements using an extended Kalman filter that adjusts confidence weights based on predicted measurement uncertainty.
All methods are evaluated on real-world data from radar and inertial sensor platforms, including both indoor and outdoor test sites with varying speed profiles. Experiments demonstrate that the learning-based radar pipeline provides consistent motion estimation and place recognition performance, while the fusion framework improves the reliability of velocity estimation compared to radar- or inertial-only baselines. This work supports the use of radar as a standalone or fused sensor for motion estimation in environments where vision or global navigation satellite systems are unavailable.
The proposed methods rely on high-resolution four-dimensional radar signal data and eliminate the need for handcrafted feature extraction or point cloud generation. Instead, neural networks are used to learn motion and spatial features directly from raw radar measurements. First, ego-motion is estimated from three-dimensional heatmap radar data using a convolutional network with attention mechanisms. Second, a representation learning framework is presented for place recognition using raw heatmaps, trained in a self-supervised manner through ego-motion as a proxy task. Third, an uncertainty-aware fusion approach integrates radar-based velocity estimates with inertial measurements using an extended Kalman filter that adjusts confidence weights based on predicted measurement uncertainty.
All methods are evaluated on real-world data from radar and inertial sensor platforms, including both indoor and outdoor test sites with varying speed profiles. Experiments demonstrate that the learning-based radar pipeline provides consistent motion estimation and place recognition performance, while the fusion framework improves the reliability of velocity estimation compared to radar- or inertial-only baselines. This work supports the use of radar as a standalone or fused sensor for motion estimation in environments where vision or global navigation satellite systems are unavailable.
Kokoelmat
- Väitöskirjat [5369]