Against the backdrop of traditional human activity recognition methods facing limitations in privacy, cost, and deployment, WiFi sensing technology based on CSI, with its strong penetration ability and wide coverage, has become an important research direction in human healthcare. In this field, deep learning has emerged as an essential tool. However, different models, due to their inherent characteristics and structures, have been limited in practical applications for human activity recognition in terms of accuracy, computational resources, and interpretability. This paper designs and implements the model named CPR-KANsformer, combining the advantages of both KAN and Transformer, featuring adaptive characteristics, high accuracy, and interpretability. Additionally, PR-KAN was introduced to reduce model parameters, and the network was expanded to achieve multi-dimensional feature extraction. In our experiments, we conducted comparative tests with four baseline models and set up ablation experiments. The results show that accuracies of 98.72% and 73.85% were achieved on both UT_HAR and Widar3.0, respectively, representing improvements of 0.61% and 2.15% compared to baseline models. This effectively demonstrates that the model designed in this paper has higher accuracy and fewer parameters in WiFi sensing human activity recognition.