Reliable indoor scene understanding requires the effective integration of complementary sensor signals. Red–green–blue–depth (RGB-D) sensing provides appearance and geometric information; however, global feature aggregation and fixed multimodal fusion can limit adaptation to varying sensor quality. In this paper, we present RegionFormer, a region-aware framework for RGB-D sensor signal processing and indoor scene classification. A shared-weight dual-stream SegFormer-B3 encoder extracts RGB and depth features, while a semantic region generator constructs 16 spatially aligned region tokens. Adaptive region feature fusion (ARFF) dynamically weights RGB, depth, and joint RGB-D features, a region-aware deep attention block enhances contextual representations, and a unified modality-adaptive transformer produces the final scene representation. RegionFormer reports 83.38% accuracy, 78.36% mean class accuracy, and 77.48% Macro F1-score on NYUD-v2, and 65.00% accuracy, 62.00% mean class accuracy, and 61.50% Macro F1-score on SUN RGB-D, with a computational cost of 11.99 giga floating-point operations per input pair on NYUD-v2. Under complete depth-sensor failure, ARFF achieves 78.60% accuracy compared with 67.83% for fixed feature concatenation, corresponding to performance reductions of 4.78 and 7.83 percentage points, respectively. Ablation studies further indicate the contribution of each architectural module. These findings highlight the potential of region-aware representation learning and adaptive multimodal fusion for robust indoor scene understanding in robotics, navigation, and intelligent automation.