In recent years, short-range wireless sensing technology based on Channel State Information (CSI) has become a popular research area. The application of deep learning has significantly improved sensing accuracy, but most existing research focuses on the selection of network models, with less attention paid to the impact of preprocessing and manual feature extraction on the results. To address this, we propose a method that combines SG filter with statistical feature extraction, aimed at enhancing the model’s efficiency in extracting features from CSI. In the preprocessing stage, we use an Savitzky-Golay (SG) filter to denoise the CSI signal, preserving the main components of the signal while filtering out noise interference. In the manual feature extraction stage, 12 statistical features are extracted from the SG-filtered CSI amplitude as input to the network. Experimental results indicate that this method performs exceptionally well in terms of sensing accuracy, computational complexity, and versatility across different networks. On the NTU-Fi public dataset, the accuracies for activity recognition and gait recognition were 99.69% and 99.12%, respectively, outperforming other existing sensing methods.