With the rapid development of intelligent transportation systems, accurate recognition of traffic states plays a pivotal role in improving traffic efficiency and ensuring travel safety. Based on video image data, a traffic state recognition model based on the Vision-KAN network was constructed to identify four core traffic states, namely traffic accidents, traffic congestion, traffic fires, and smooth traffic. To verify the effectiveness of this model, the basic models of residual networks (ResNet), convolutional neural networks (CNN) and vision transformer (ViT) were selected for comparison. The experimental results show that the Vision-KAN model outperforms the comparison models in all aspects and demonstrates a high-precision recognition ability for traffic states in video images. It can provide strong support for intelligent traffic management and control as well as emergency response, and enhance road traffic efficiency and security guarantee capabilities.