This paper addresses the challenge of integrating multimodal data for enhancing service quality at railroad passenger stations. Current railroad data platforms hold vast amounts of multimodal data, but the lack of effective cross-modal feature fusion and analysis techniques limits their utilization. We propose a joint analysis framework that integrates video, image, and audio data to improve emotion recognition at railroad service desks. By leveraging a deep learning-based emotion recognition network, our model extracts and fuses features from body language, scene context, and speech to accurately identify passenger emotions. This capability enables real-time, personalized responses, enhancing passenger satisfaction and mitigating potential conflicts. Experimental results demonstrate that the proposed multimodal approach outperforms unimodal and bimodal models in emotion recognition accuracy, recall, and F1 score, proving its effectiveness in real-world scenarios. This research not only enhance the level of intelligence in railway consulting services, but also represent a significant step for the railway industry to adapt to digital transformation and meet the diverse needs of passengers.