AIDE-SLAM: Anomaly-Handling and IMU-Enhanced SLAM

This paper presents AIDE-SLAM, a robust visual-inertial SLAM framework integrating hierarchical anomaly handling and IMU-enhanced optimization for dynamic environments. Addressing traditional VSLAM vulnerabilities to illumination changes, motion blur, and sensor anomalies, we propose a cascaded anomaly detector combining photometric screening (KL divergence, Laplacian sharpness) with a bidirectional LSTM network, achieving 97.2% precision in 5 ms. A confidence-driven optimizer dynamically adjusts sensor fusion weights (λ=0.5(1-α)e^(-γΔt)) across three modes: IMU-dominant (α≥0.7), hybrid (0.4≤α<0.7), and vision-priority (α<0.4), reducing IMU-only error by 25% on EuRoC datasets. Experiments show 37.5% lower ATE RMSE (0.049m vs. 0.075m) in high-dynamic sequences and 81.6% improvement under synthetic anomalies (e.g., 40% occlusion). Keyframe spatial consistency enables rapid recovery, achieving 5.2× faster mode transitions. The system maintains sub-decimeter accuracy (0.033m avg. ATE) in degraded conditions, demonstrating practicality for real-world robotics and AR applications.