Sparse representation is a powerful tool in signal denoising, and visual evoked potentials (VEPs) have been proven to have strong sparsity over an appropriate dictionary. Inspired by this idea, we present in this paper a novel sparse representation-based approach to solving the VEP extraction problem. The extraction process is performed in three stages. First, instead of using the mixed signals containing the electroencephalogram (EEG) and VEPs, we utilise an EEG from a previous trial, which did not contain VEPs, to identify the parameters of the EEG autoregressive (AR) model. Second, instead of the moving average (MA) model, sparse representation is used to model the VEPs in the autoregressive-moving average (ARMA) model. Finally, we calculate the sparse coefficients and derive VEPs by using the AR model. Next, we tested the performance of the proposed algorithm with synthetic and real data, after which we compared the results with that of an AR model with exogenous input modelling and a mixed overcomplete dictionary-based sparse component decomposition method. Utilising the synthetic data, the algorithms are then employed to estimate the latencies of P100 of the VEPs corrupted by added simulated EEG at different signal-to-noise ratio (SNR) values. The validations demonstrate that our method can well preserve the details of the VEPs for latency estimation, even in low SNR environments.