Artifacts that contaminate electroencephalography (EEG) signals make it difficult to analyze EEG. The aim of this study was to removal artifacts on EEG, especially those caused by motion, to measure EEG in unconstrained situations. In a previous study, head movements were detected by an accelerometer, and motion artifact components were separated from the recorded EEG by independent component analysis (ICA). This method is effective for reducing the effect of artifacts, but has a risk that EEG components are also removed. In this paper, we introduce an improved artifact removal method based on ICA and filtering. EEG were decomposed by ICA, and a Pearson's correlation coefficient was calculated between each independent component and each hybrid accelerometer value to distinguish artifact components. Artifact components were then high-pass filtered. In this study, subjects were instructed to move their heads randomly, while keeping their eyes closed. The previous method was adapted using 1, 2, 3, 5 and 10 s to find the most suitable epoch to minimize the mean absolute amplitude of the cleaned EEG. Then, using this epoch, the proposed method was compared with the previous method by frequency analysis. Low frequency power (0.1–3 Hz) was normalized to unity because most power caused by motion artifacts exists in the low power band. If the normalized theta (4–8 Hz), alpha (8–13 Hz) and beta (13–40 Hz) powers of cleaned EEG are higher than that of raw EEG, this indicates that the effect of motion artifacts is small and EEG components are retained. The results obtained from theta and alpha power comparison showed that the proposed method performed better than the previous method. This result suggests that the proposed artifact removal method is more effective to reduce the effect of artifacts while retaining the EEG components.