In the non-contact heart rate detection method based on ballistocardiogram， the actual shape of ballistocardiogram signals is easily covered up during notable body movements. To address the obstruction caused by invalid signals in locating the heartbeat point， a waveform compensation model for notable movement segments is proposed， which combines phase space reconstruction with RBF neural network. Firstly， the improved C-C method is used to select the appropriate reconstruction parameters. Then， the network topology is determined by dynamic k-means clustering. Transform the time series before the movement into phase points in reconstructed space， and feed them into the model as learning samples. Finally， the single-step recursive prediction of invalid signal segment is realized. Experimental results show that the prediction model has good accuracy and it can reduce the influence of irregular noise in the original signal. After model modification， the mean error of beat by beat cardiac cycle is 1.27% and the mean absolute error is 8.9 ms， effectively avoiding the misjudgment of heartbeat events.