<p>Emotion recognition is pivotal in advancing human-computer interaction, with transformative applications in education, healthcare, and social analysis. Traditional models predominantly rely on facial expressions, often neglecting the critical roles of contextual and bodily cues. In this study, we propose mCFA, a novel context-aware emotion recognition model that integrates facial, body, and contextual cues through a two-level attention mechanism. First, cross-attention modules dynamically learn inter-modality interactions; second, adaptive fusion assigns weights to each modality based on its contribution. Evaluated on the EMOTIC dataset, mCFA achieves a mean Average Precision (mAP) of 28.77%, and an Accuracy of 85.53% on the CAER-S dataset, outperforming several state-of-the-art models. Our approach demonstrates robust performance, particularly in the presence of occlusion or missing visual cues, underscoring the effectiveness of attention-driven fusion for emotion understanding.</p>