K
KITT
Text
HybridTrack: Learnable Kalman Filter for 3D Multi-Object Tracking
This Idea covers a hybrid 3D multi-object tracking method that replaces the hand-crafted parts of a Kalman-filter tracker with learned ones inside a tracking-by-detection pipeline. Rather than assuming a fixed motion model and manually tuned process and measurement noise covariances, the system uses neural modules to predict the state-transition residual, the Kalman gain, and the noise statistics from data, so it adapts across traffic scenarios without per-scene tuning. It retains the recursive predict-update structure -- and thus the interpretability and real-time efficiency -- of a classical filter while gaining the generalization of deep learning. Learners see how a data-driven Kalman filter is trained and evaluated (KITTI, HOTA) and why hybridizing model-based filtering with learning beats both purely model-based and fully end-to-end trackers.