A Decision Stabilization Method for Domain Transition Regions Using Core Feature-selection-based Time-series Classification and EMA and HMM Stabilization

Mobile robots operating in disaster environments frequently transition between indoor and outdoor spaces, where abrupt changes in illumination, background appearance, and structural patterns occur. These domain transition regions often lead to unstable frame-level environmental perception. This study analyzes the visual uncertainty and decision fluctuations observed in such regions and proposes a sequence-based approach for stable domain transition recognition. A video dataset containing domain transitions is constructed, and 30 visual features—including classification probabilities, entropy, brightness variation, edge density, structural cues, and motion information—are extracted. SHAP-based feature importance analysis is employed to identify the top 14 key features that most significantly contribute to transition region discrimination. Using these selected features, sequence-based classification models based on BiLSTM, TCN, and Tiny Transformer are trained. In addition, EMA- and HMM-based stabilization techniques are applied to suppress decision fluctuations in transition regions. The experimental results show that sequence-based models outperform frame-level classification in transition detection accuracy and decision stability. In particular, the combination of TCN and HMM achieves the best overall performance. These findings demonstrate that integrating sequence modeling with stabilization techniques is an effective strategy for recognizing domain transition regions characterized by gradual visual changes, thereby improving the real-time environmental perception stability of disaster response robots