The KITTI Vision Benchmark Suite

This paper presents ICCA-MOT, a real-time 3D
multi-object tracking framework that combines
innovation-consistent adaptive Kalman filtering
(IC--AKF), confidence-aware cascaded association
(CACA), and a projection-based visual validation.
IC--AKF jointly shapes and scales the process and
measurement covariances from innovation
statistics, restoring near-white innovations and
numerical stability under time-varying noise. CACA
integrates 3D GIoU with motion and shape penalties
under uncertainty-aware weights, and prioritizes
reliable hypotheses via confidence staging. The
visual check suppresses visually unsupported LiDAR
artifacts with negligible overhead while
preserving LiDAR-only coverage.

\

@article{2025ICCA,
title={Innovation-Consistent, Confidence-Aware
Real-Time 3D Multi-Object Tracking},
year={2025},
note={under review}
}