CATER: Causal Adaptive Tracking with Environmental Reasoning

Robust state estimation for tracking multiple targets remains challenging under varying environmental conditions. We introduce CATER (Causal Adaptive Tracking with Environmental Reasoning), a framework that integrates dynamic causal modeling with hierarchical intervention detection and adaptive variational filtering. CATER explicitly captures time-varying relationships between environmental factors and measurement quality, enabling principled adaptation of the state estimator to changing conditions. Our analysis indicates favorable identifiability and stability properties. Experiments in both controlled simulations and realistic autonomous driving scenarios (including CARLA and nuScenes) demonstrate CATER significantly outperforms state-of-the-art tracking components when integrated into standard MOT pipelines. This results in substantial improvements in state estimation accuracy reflected in MOT metrics (e.g., 6.74-7.91 percentage points gain in MOTA on CARLA) and consistency (e.g., 34% fewer ID switches on CARLA; lowest IDS on nuScenes adverse subsets). The framework shows particular advantages in challenging weather conditions (fog, rain, night) while maintaining real-time performance (e.g., 208 FPS in CARLA evaluation), addressing a fundamental limitation in existing approaches that often treat dynamic environmental factors affecting state estimation merely as static parameters or unstructured noise.