Freezing of gait (FOG), a paroxysmal motor block, happens typically in the latter stages of Parkinson's disease (PD), and it is a main cause of falls in PD. Nowadays, clinical evaluation of FOG is still dependent on the knowledge of different expert raters. Hence, the assessment is often subjective and self-contradictory. An objective management of FOG is very necessary, but is also extremely difficult due to its sudden and transient property. This research intends to investigate the application of a machine learning method to detect the FOG events by using gait data recorded by wearable accelerometers. Total ten features were extracted from the temporal and the spectral domain of the acceleration signal. When using a sliding window with a length of 4 seconds to divide the acceleration time series sampled from 10 PD patients, a total of 3463 FOG instances and 32187 non-FOG instances were obtained. A classification algorithm that combined boosting algorithm and random sampling technique was utilized for the classification of minor FOG samples and major non-FOG samples. To obtain an optimal feature subset, we also implemented a feature selection algorithm that was based on the normalized mutual information. The best results obtained in the subject-dependent experiments were an average sensitivity of 99.70% and specificity of 99.96%.