Sundown Syndrome (SS) is a condition characterized by cognitive, emotional, and behavioral disturbances, primarily affecting individuals with dementia in the late afternoon and evening. Despite being long recognized, research on SS remains limited due to the lack of standardized diagnostic criteria and the limited availability of structured multimodal datasets. To address this gap, we present SunVid, a curated online video database specifically designed for SS research. This dataset includes 74 SS-related videos from online platforms and 405 annotated segments labeled with SS states, affective states, symptoms, and context. To assess the effectiveness of SunVid, we perform three evaluation experiments using facial features, Emonet-extracted emotional features, and body gesture features, employing state-of-the-art benchmark models to establish baseline performance. For facial feature analysis, the Swin Transformer model achieves the highest accuracy of 69.03% in SS vs. non-SS classification. Emonet-extracted emotional features, including emotion category, valence, and arousal, produce slightly better results with an accuracy of 72.85%. Body gesture analysis across different models further confirms that movement patterns contribute to SS identification, although with lower predictive accuracy than facial features. Despite limitations such as class imbalance and variations in video quality, our findings demonstrate the feasibility of AI-assisted SS detection. This study also highlights the potential of video data as a novel avenue for SS analysis and underscores the promise of AI-driven methods in supporting SS detection and intervention.