To address the challenges of privacy protection in recognizing unsafe behaviors such as falls and low cross-environment recognition rates, this paper proposes a behavior recognition framework, Single-antenna Cross-environment Stable Human Activity Feature Extraction and Recognition Framework (SSRF), based on Channel State Information(CSI), optimized from the existing ReWiS model. By collecting data on five types of elderly behaviors(such as falls, no action, etc.) from different environments, the CSI signals are normalized, followed by Singular Value Decomposition(SVD) and Pearson correlation coefficient calculation to generate labeled CSI data samples, which are then fed into the ProtoNet model for classification. Compared to ReWiS, SSRF significantly reduces the number of parameters(from 111 936 to 37 392) and accelerates both training and testing speed, with total training time reduced from 33.12 s to 26.8 s, and per-sample testing time reduced from 0.000 149 s to 0.000 104 s. In the four-category task of a public dataset and the five-category task of a custom dataset, SSRF achieves average cross-environment recognition accuracies of 89% and 85%, respectively, with 95% accuracy for fall detection. Experimental results show that SSRF maintains high generalization performance while significantly improving the efficiency.