Towards Adaptive Video Stabilization: A Robustness Benchmark of IMU Predictors for Cascaded Routing

Ilya Afanasyev

Keywords: Trustworthy AI, Explainable AI, Shapley Values, Bayesian Neural Networks, Edge Computing, Cascaded Inference, Anomaly Detection

TL;DR: We benchmark 12 early-stage IMU predictors for cascaded video stabilization, demonstrating that a Bayesian 1D-CNN (0.38 ms) with epistemic uncertainty bounds and SHAP-guided kinematic routing ensures trustworthy edge-AI inference.

Abstract: Deploying energy-intensive computer vision algorithms on edge devices requires architectural shifts to minimize computational overhead. We propose a cascaded inference paradigm for adaptive video stabilization, utilizing a lightweight Inertial Measurement Unit (IMU)-based model to route motion anomalies to heavy verification solvers. To establish a rigorous foundation for this cascaded router, we benchmark 12 multi-paradigm models aligned with Trustworthy AI principles: evaluating predictive accuracy, strict single-sample sequential latency, and out-of-distribution (OOD) robustness under severe sensor noise. Additionally, Game-Theoretic SHAP analysis extracts physical kinematics, quantifying translational (accelerometer) versus rotational (gyroscope) contributions to inform downstream stabilization strategies (OIS vs. EIS). Our results reveal that standard batch-latency metrics mask significant API overheads in gradient boosting frameworks. Consequently, CatBoost emerges as the optimal deterministic router (0.44 ms), minimizing serialization costs. Crucially, our evaluation of the probabilistic Bayesian 1D-CNN exposes a vulnerability to confident misclassification under severe noise, demonstrating that epistemic uncertainty alone is insufficient for fail-safe routing. However, achieving sub-millisecond execution (0.38 ms) and near-perfect accuracy, the Bayesian architecture provides mathematically rigorous epistemic bounds. This formally justifies orthogonal Isolation Forest integration for OOD interception. Ultimately, by coupling high-speed predictive routing, mathematically rigorous safety bounds, and game-theoretic XAI, this work advances the formal basis for designing reliable hybrid edge-AI systems.

Submission Number: 140