Aiming at the problems of excessive computational complexity and large number of parameters in existing radar fall detection methods, this paper proposes a lightweight adaptive convolutional neural network, LACNN.Firstly, micro-Doppler features are extracted from the human activity echo signals acquired by the FMCW radar. Secondly, Initial extraction of features using an improved lightweight ShuffleNet network. Then, the features are refined using a lightweight channel-space attention module with an efficient asymmetric convolutional kernel parallel multi-scale feature extraction module. To improve the generalisation ability of the model, a convolutional batch normalization AconC module is embedded in the model. Finally, the fused features are fed into the fully connected layer for detection. Comparison results with other network models show that the proposed model achieves an F1 score of 99.33%, which is an improvement of 0.61%~4.10%, while maintaining a lower computational cost, with FLOPs of only 1.047 M and model parametric quantities of only 69.09 M.