MOTCues: 3D Multi-Object Tracking with Birth Prior and Shape Description Informed by Point Cloud Cues
Hanyeol Lee, Yeongkwon Choe, Taeyoon Kim, Chan Gook Park
AI summary
Problem
Traditional model-based multi-object trackers struggle with ambiguous object birth/death events, detector errors, and ID switches in complex scenarios due to reliance on uniform birth priors and simple spatial distance metrics.
Approach
MOTCues models the object birth intensity using Gaussian mixture fitting on segmented point clouds and incorporates bounding-box-centric shape descriptors into the track-to-measurement association cost function within a Poisson multi-Bernoulli filter.
Key results
- Formulates a conjugate Gaussian mixture birth prior from raw point clouds to improve new-object initialization.
- Introduces a Bbox-centric shape descriptor mapped to the probabilistic domain to enhance hypothesis management.
- Demonstrates reduced ID switches and superior tracking performance compared to baseline model-based trackers on KITTI and nuScenes benchmarks.
- Maintains computational efficiency by preserving the conjugate family of the PMB filter without relying on heavy neural networks.
Why it matters
Provides a mathematically interpretable, computationally efficient tracking framework that enhances safety and reliability for autonomous vehicles and mobile robots operating in complex environments.
Abstract
Reliable multi-object tracking (MOT) is essential for autonomous systems but remains challenging due to am- biguous object characteristics such as birth, death, and motion models, as well as detector errors including false detections and missed objects. Random finite set (RFS) theory provides a rigorous mathematical foundation that enables the formulation of fundamental uncertainties in object estimation under the Bayesian framework. We propose MOTCues, a MOT algorithm built on the RFS-based Poisson multi-Bernoulli filter, which integrates informative components derived from point cloud cues into the estimator as a tailored formulation. The object birth intensity function is modeled with a Gaussian mixture distribution for effective initialization of new-born objects, while object shape information is captured by constructing bounding box-centric descriptors to enhance hypothesis management. Evaluations on the KITTI dataset and the nuScenes benchmark demonstrate that integrating point cloud cues improves tracking performance by reducing ID switches, achieving superior results compared to baseline model-based trackers in real-world object tracking scenarios.