Event-related potential (ERP) plays a very important role in the field of human brain activities research. It is a practical method for measuring the brain functions. By now, the traditional methods remained in extracting of ERP are that rely on accumulative averaging techniques, which getting in a totally averaging result. In practice, however, it is obviously that the ERPs are not identical with each other in response for a number of repeated stimuli, neither in signal pattern nor response time. So that extracting ERP from a single trial is the goal of investigators in pursuit of. That is a different task, although some worthy works had been reported. A novel method is presented in this paper, which can extract single trial ERP by means of higher order cumulant (HOC) followed by cepstrum technique. Based on the theory of HOC, it can deal with additive noise very well, regardless the noise is white or not. For a single-trial ERP signal measured in strong background noise, the complex cepstrum of higher order cumulants of the signal is calculated firstly, and then the original ERP is reconstructed. The experiment shows that this method has a better performance in reconstructing single-trial ERP in the case of lower signal to noise ratio.