In this study, we attempted to characterize body motion pattern in patients with Parkinson's disease during walking. Body motion was recorded from 16 patients with Parkinson's disease and 10 healthy elderly subjects by using a tri-axial accelerometer without any constraints. To assess body motion pattern, time-frequency patterns of the acceleration signal were estimated using a matching pursuit algorithm. The time-frequency patterns were totally different between the healthy elderly subjects and patients with Parkinson's disease. We classified the time-frequency patterns into four distinct patterns based on their time durations: duration patterns I(D-I:< 0.15s), II(D-II:0.15-0.5s), III(D-III:0.5-2.0s)and IV(D-IV:≥ 2.0s). The energy of the D-II patterns representing the irregular activities and the energy of the D-III patterns representing the short-term regular(rhythmic)increased in the patients with Parkinson's disease compared to those of the healthy elderly subjects. On the other hand, the energy of the D-IV patterns representing the long-term regular(rhythmic)activities significantly decreased in the patients with Parkinson's disease. Furthermore, these features were seen more clearly in severe case patients than in mild case patients. We concluded that these differences are likely due to a lack of ability to control normal and smooth movement in Parkinson's disease. Therefore, the time-frequency patterns of acceleration signal can be used to assess gait disorder in patients with Parkinson's disease.