PiVoT: A Training-Free Tracker for Large-Scale Radar Point Clouds in Heavy Clutter

Introduction

Radar perception often has to operate under conditions that are unfriendly to modern learning pipelines: noisy point clouds, severe clutter, changing object counts, and limited labeled data. Deep detectors can perform well when trained on suitable datasets, but collecting and maintaining those datasets across sensors and environments can be difficult. Classical Bayesian trackers, on the other hand, are attractive because they can work without training, yet they often struggle when clutter becomes dense or the number of objects grows.

PiVoT, a new arXiv paper, proposes a variational solution for this gap. It is designed as a fast, training-free multi-object detector and tracker for both positional and Doppler measurements, aiming to handle full-resolution radar point clouds in real time.

Key ideas

• Joint detection and tracking without a separate detector: PiVoT does not depend on external clustering or a pretrained detection network. Instead, it performs end-to-end inference directly from noisy measurements.
• A richer multi-object state model: The method jointly estimates object states, shapes, existence probabilities, data association, and measurement rates. This allows the tracker to reason about both whether an object is present and how incoming points should be assigned.
• Designed for heavy clutter: The paper focuses on scenes where clutter may be visually inseparable from true objects. This is one of the hardest settings for radar point-cloud tracking, because false measurements can look similar to real object returns.
• Variational inference for scalability: PiVoT introduces several efficiency-oriented updates, including theoretically justified birth pruning, reductions from quadratic to linear complexity for exact updates, and a computationally efficient Doppler Poisson model.
• Automotive radar relevance: The experiments described in the abstract show real-time operation on full-scale modern automotive radar datasets. The authors also state that PiVoT reaches performance comparable to a deep-learning detection benchmark while remaining training-free.

Why it matters

The main significance of PiVoT is not simply that it improves a tracker, but that it strengthens the case for probabilistic, training-free perception in radar systems. Many practical radar deployments do not have abundant labeled data, and even when labels are available, distribution shifts between sensors, environments, and traffic patterns can be substantial.

By combining Bayesian modeling with more efficient variational inference, PiVoT aims to preserve the interpretability and data efficiency of classical tracking while addressing the runtime and scale limitations that have kept such methods from broader use in modern high-resolution radar pipelines.

Its support for Doppler measurements is also important. Modern automotive radar provides velocity-related information that can be highly valuable, but using it efficiently at full resolution is computationally demanding. PiVoT’s Doppler Poisson modeling is presented as a way to make that information usable without relying on a separate learned detector.

The abstract does not provide enough detail to judge generalization across all datasets or deployment conditions. Still, the direction is notable: high-performance radar perception may not have to be purely deep-learning-based. Carefully engineered probabilistic inference remains a serious option, especially where robustness, interpretability, and limited training data are central constraints.

Source: arXiv