PiVoT delivers real-time multi-object tracking under heavy clutter without training

Multi-object detection and tracking from noisy point clouds remains a critical challenge in radar applications like autonomous driving, where heavy clutter and large object populations overwhelm traditional approaches. Current Bayesian trackers using Poisson measurement models are training-free but struggle with accuracy and efficiency. Enter PiVoT, a new variational inference solution developed by researchers at the University of Edinburgh and others. PiVoT jointly infers object states, shapes, existence probabilities, data association, and measurement rates end-to-end—no external clustering, detection, or training required. Its efficiency comes from theoretically justified birth pruning, a quadratic-to-linear complexity reduction for exact updates, and a computationally efficient Doppler Poisson model.

In experiments, PiVoT substantially outperformed existing Bayesian trackers in challenging scenes with severe clutter and full-resolution Doppler point clouds. It demonstrated exceptional scalability, handling up to 1,000 objects simultaneously, and operated in real-time on full-scale modern automotive radar datasets. Remarkably, as a training-free joint detector and tracker, PiVoT achieved performance comparable to deep learning detection benchmarks. This combination of speed, accuracy, and robustness makes PiVoT a promising solution for real-world applications such as autonomous vehicles, surveillance, and robotics—where reliable multi-object tracking under clutter is essential.