To improve the performance of AI applications, large-scale models have received increasing attention due to their excellent generalization ability. However, the training and transmission of large-scale models impose significant computational and communication burdens on resource-constrained edge nodes, and the exchange of complete models may violate model privacy. To reduce the burden on nodes and protect model privacy, split federated learning (SFL), which integrates data parallelism and model parallelism, is proposed. However, in addition to the challenge of resource limitations, SFL also faces two other critical challenges in edge computing (EC) systems, i.e., statistical heterogeneity and system heterogeneity. To address these challenges, this study proposes a novel, efficient SFL framework based on feature compression and feature fusion, termed SplitCP, by incorporating the characteristics of SFL. Specifically, feature compression aims to reduce the communication resource consumption of nodes and improve training efficiency by assigning appropriate compression ratios to heterogeneous nodes. Feature fusion aims to merge the features of nodes into a mixed feature sequence, which is approximately equivalent to the features of independent and identically distributed (IID) data, thus overcoming the challenge of statistical heterogeneity and improving model accuracy. Moreover, by analyzing the coupling relationship between adaptive feature compression and feature fusion, SplitCP explores how to jointly optimize them to improve the model training performance of SFL. Extensive experiments are conducted on a physical platform built with 80 NVIDIA Jetson edge devices, and the experimental results show that, compared with the baselines, SplitCP can improve the final model accuracy by 6.46% to 26.15% while achieving training speedups of approximately 1.31× to 3.51×.