In this paper, we propose a CLSA human activity recognition (HAR) method based on channel state information (CSI) to address the challenges of environmental noise and feature overfitting. The method adopts a hybrid deep learning architecture that combines a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network and an attention mechanism. Raw CSI data are first denoised using discrete wavelet transform (DWT) to preserve activity-related information while suppressing random noise. The preprocessed data are then passed through the CNN layer to extract spatial features using a two-layer Conv1D structure. These spatial features are further processed by a two-layer stacked LSTM network to model temporal patterns. The output of the final LSTM layer is then fed into the attention layer, where weights are dynamically assigned to emphasize the contextual relevance of key activity segments. Finally, the classification layer is used to categorize the activities. The experimental results show that the CSI-based human body recognition method proposed in this paper outperforms models such as LSTM and CNN and some recognition methods on real datasets. The average recognition accuracies of up to 90.09% and 89.64% are achieved in two different environments, respectively, providing an effective method for activity recognition.