Doppler-aided Position-Velocity Tracking of Hidden Objects using mmWave Radar
Authors: Yerin Cheon, Kush Garg, Akshat Dave
Organizations: Stony Brook University, 100 Nicolls Rd, Stony Brook, NY 11794, United States. · Delhi Technological University, 110042, New Delhi, India.
Safe indoor operation requires autonomous tracking of objects that are not directly visible to the robot. Optical non-line-of-sight (NLoS) methods depend on ambient lighting and scene geometry or on weak multi-bounce returns that need long acquisitions. Millimeter-wave (mmWave) radar offers an attractive alternative: walls act as reflectors for mmWave signals, and frequency-modulated continuous-wave (FMCW) radars enable fast, practical measurement of range, angle, and velocity. However, the low resolution of these radars leads to spatially ambiguous observations. The Doppler information, although particularly sensitive to target motion, remains underexplored in indoor NLoS localization: existing methods emphasize spatial representations and either collapse or discard the Doppler dimension before learning. To leverage the complete Range-Azimuth-Doppler (RAD) information, we acquire the first indoor NLoS radar dataset that pairs sequential RAD measurements with synchronized position and velocity ground truth under controlled variations in motion: 576 sequences totaling 720,856 frames across 12 trajectories of a mobile robot moving along varying speeds, starting positions, and material configurations. We then formulate NLoS tracking as sequential position-velocity estimation rather than frame-wise localization: a shared 3-D encoder processes the full RAD representation of consecutive frames, separate recurrent branches estimate position and velocity, and post-inference refinement combines the two into a motion-consistent trajectory. Our method achieves 8.95cm position and 1.75cm/s velocity errors on held-out trajectories and generalizes to unseen scenarios involving a moving human target. Together, our dataset and tracker establish a benchmark for tracking hidden moving targets with a low-cost radar.
Figures & tables
Fig. 2 : Materials and trajectories in our dataset acquisition . Our dataset includes four target configurations and 12 trajectory patterns.
Fig. 3 : Our NLoS position-velocity tracking pipeline. Input RAD frame comprises 5 channels for each time step: normalized power, validity mask, and positional embeddings for RAD dimensions. A shared 3D residual CNN encodes the RAD frame into a 48-D feature. Independent position and velocity branches add a gated recurrent unit (GRU)- based temporal correction. The GRU hidden state is updated through eight consecutive RAD frames. The resulting position and velocity estimates are then refined for motion consistency.
Fig. 4 : Mean predicted and ground-truth position and velocity magnitudes over the validation set. Values at each time step are averaged across all validation sequences. The predictions closely follow the ground truth over time, showing that the position estimate does not drift and the velocity estimate remains accurate with time.
Speed
ex
ey
evx
evy
(m/s)
(cm)
(cm)
(cm/s)
(cm/s)
0.25
3.69
3.59
0.79
0.80
0.75
6.48
6.34
1.22
1.15
1.00
6.65
6.37
1.35
1.28
TABLE I : Mean absolute error at different target speeds. Target speed is measured using Kobuki wheel odometry and reported in m/s.
Fig. 5 : Qualitative tracking results on six distinct real-world trajectories. For each trajectory, the left panel shows the ground-truth position and color-coded velocity, while the right panel shows filtered position estimates colored by the velocity estimated by our pipeline. The results qualitatively show that the proposed method accurately estimates both position and velocity across diverse real world trajectories.
Fig. 6 : Trajectory-wise mean absolute error across different material configurations. Position errors ( ex , ey ) are reported in cm and velocity errors ( evx , evy ) in cm/s; lighter cells denote lower error. Each cell is averaged over three target speeds and four starting positions, and all errors are reported after applying motion-consistent trajectory refinement.
Configuration
Position MAE (cm)
Velocity MAE (cm/s)
Full model
8.95
1.75
w/o GRU Temporal Memory
11.16
2.10
w/o Azimuth
40.82
3.39
w/o Doppler
71.22
8.35
w/o Range
46.94
4.04
TABLE II : Ablation study of radar input dimensions and temporal memory. To ablate azimuth, Doppler, or range, we average the power and validity mask along that dimension, broadcast the mean back to the original shape, and zero the corresponding coordinate channel. To remove temporal memory, we reset both GRU hidden states before each frame.
Fig. 7 : Comparison with similar architecture but using an internal learned filter and coupled position-velocity updates. The coupled model exhibits position drift during stationary intervals, whereas our independent estimation followed by temporal refinement reduces drift. Across the full held-out N-validation set, the coupled model achieves mean position and velocity errors of 21.79 cm and 3.27 cm/s, compared with 8.95 cm and 1.75 cm/s for ours.
Fig. 8 : Evaluation of generalization across unseen targets and trajectories. A walking person (training used only the TurtleBot) on a trajectory absent from training; Our model approximately recovers the position and velocity trajectory, demonstrating good out-of-distribution generalization.
mmWave radars are robust to darkness and occlusions such as dust and smoke, and can directly constrain ego-velocity from a single frame via Doppler measurements, making them attractive sensors for odometry in visually denied conditions. However, almost all existing radar-inertial odometry systems rely on lossy radar point clouds that are highly sparse, generally concentrated in a narrow angular band, and aliased at high speeds. We propose to instead estimate ego-velocity directly from unfiltered mmWave I/Q signals. Taking advantage of a foundation model for 4D radar spectrum, we develop a system that integrates uncertainty-aware velocity predictions with IMU measurements using an uncertainty-weighted sliding-window pose graph to accurately compute odometry even when provided radar data with aliasing or an unfavorable field of view. Evaluated on public benchmark datasets, our system, UNRIO, attains the lowest relative pose error on the majority of sequences across held-out environments, despite differing chirp configurations, motion patterns, and platforms with its strongest gains coming where point clouds fail most: lateral motion in IQ1M and heavily aliased Doppler in ColoRadar.
Jui-Te Huang, Tianshu Huang, Anthony Rowe +1
Carnegie Mellon University, Pittsburgh, PA 15213, USA
Spinning frequency-modulated continuous-wave (FMCW) radars have been gaining popularity in autonomous vehicle perception on account of their robustness to adverse weather conditions and 360° field of view. Recently, scanning radars have also been shown capable of generating per-azimuth Doppler velocity. In this paper, we investigate whether these Doppler velocity measurements improve spinning radar vehicle detection and tracking performance. For detection, we estimate the ego motion and use it to undo the Doppler range distortion of the radar image before passing it to a network. For tracking, we propose a new way to estimate a per-vehicle velocity and use it as a prior for the tracker's motion model. Since Doppler-enabled spinning radar data is not available in any dataset with ground-truth dynamic object labels, our first contribution is an automatic labelling pipeline that uses an ensemble of fine-tuned off-the-shelf lidar detectors to label all 643 km of the Boreas Road Trip dataset. We then transfer detections to radar, and use over 250 km of vehicle-dense sequences as ground-truth training data. By training and evaluating two state-of-the-art detectors, we show that Doppler undistortion can improve detection accuracy by up to 2.37 points on mean average precision. Furthermore, we show that the Doppler velocity prior can improve tracking accuracy by 13.68 points on multi-object tracking accuracy (MOTA) versus the zero-velocity initialization baseline, while achieving 99.7% of the MOTA obtained using ground-truth velocities as the prior.
Sensing ego-velocity estimation is fundamental to state estimation in visually degraded environments, where camera- and LiDAR-based pipelines can become unreliable. Millimetre-wave radar is well suited to these conditions because it provides direct Doppler velocity sensing and remains robust to poor illumination, textureless scenes, and airborne particulates. However, conventional radar ego-velocity pipelines typically apply constant false alarm rate (CFAR) thresholding to convert dense radar spectra into sparse point clouds, prematurely discarding sub-threshold returns that may still retain useful Doppler motion cues. We present Dense Soft Weighting, an analytic radar front-end that maps every range-Doppler cell to a continuous confidence metric rather than enforcing a binary detection threshold. Ego-velocity is then estimated using a deterministic robust weighted least-squares formulation, while the same weighted measurements provide a closed-form, measurement-derived velocity covariance for integration with a shared inertial back-end. The method requires no platform-specific training data or learning-based uncertainty model, supporting transfer across single-chip radar configurations. Across two public datasets and one self-collected dataset, Dense Soft Weighting reduces mean absolute pose error by 31-45% relative to the strongest CFAR point-cloud baseline under an identical inertial back-end, while running in real time on embedded hardware.