To solve the problems of insufficient time-frequency feature representation ability and insufficient neural network feature learning in the existing radar human behavior recognition, a human behavior recognition method based on radar time-frequency feature extraction and CBAM-MFResNet is proposed. In the time-frequency feature extraction section, radar echo signals are processed, the distance window function is used to constrain the spectral energy diffusion problem in the behavioral signal, and along the slow time dimension, an adaptive wavelet threshold-Chebyshev window function co-processing mechanism is constructed to suppress clutter interference. Micro-Doppler time-frequency diagrams are obtained by time-frequency analysis. In the network model building section, a CBAM-MFResNet model for behavior recognition is constructed, the lightweight convolutional attention mechanism is introduced into the residual neural network to enhance the representation of key features; and an efficient parallel multi-scale feature learning module was designed to learn diverse feature information to reflect the feature differences of different behaviors to the greatest extent. Finally, the fused features are input into the fully connected layer for classification. Experimental results show that the proposed model and clutter filtering algorithm can effectively improve the accuracy of the recognition system, and the average recognition accuracy of different human behaviors reaches 98%.