MIMOT: Mamba-based inter-trajectory interaction for multi-object tracking.

Abstract

Online multi-object tracking (MOT) requires reliable motion prediction when detections are missing or targets occlude one another. Existing drop-in state-space model (SSM) motion predictors mainly derive each update from the target's own history, which limits their use of nearby motion context. We propose MIMOT, an online tracking-by-detection method that replaces the Kalman filter with a geometry-aware Mamba motion predictor while retaining YOLOX detection and ByteTrack-style association without appearance cues. MIMOT first encodes contiguous and temporally strided track histories and fuses their complementary temporal evidence. It then conditions nearby trajectories on relative geometry and aggregates them through a distance-ordered bidirectional Mamba scan whose aggregation cost is linear in the selected neighborhood size. A codebook-conditioned SSM decoder subsequently predicts residual box updates using input-dependent mixtures of transition parameters. MIMOT achieves 58.1, 74.2, and 62.1 HOTA on the official DanceTrack, SportsMOT, and MOT17 test sets, respectively. It ranks first among the evaluated learned-motion trackers on DanceTrack and MOT17 and remains competitive on SportsMOT. Controlled ablations further show that temporal fusion, geometry-conditioned interaction, and codebook decoding mainly improve association accuracy.