Over the past few decades, transportation infrastructure has expanded significantly. However, traffic issues have proliferated due to urban population growth, necessitating increased use of various transport modes. Despite existing traffic prediction models having made progress, they still face challenges like high data dependence, model complexity, and real-time performance issues. This prompts a reassessment of traffic flow prediction methodologies. To address this issue, we proposed a real-time monitoring optimization model based on deep learning. First, traffic flow parameters (such as traffic flow density, flow, speed, etc.) are extracted from video in real-time using YOLOv5 and Deep SORT technology and combined with the basic traffic flow model to analyze time changes in these parameters. Secondly, we use a transformer model for traffic flow prediction. Following this, a real-time traffic congestion warning system is proposed based on the predicted traffic flow data, and an emergency lane activation decision-making model is established by designing algorithms to improve road traffic efficiency and emergency response capabilities. Experimental results show that the method proposed in this paper has significant advantages in the accuracy, real-time performance, and decision support capabilities of traffic flow prediction, providing an effective solution for modern urban traffic management.