Freezing of gait (FoG) is a typical movement disorder in Parkinson’s disease (PD) that severely impacts on patient’s quality of life. FoG detection based on electromyographic (EMG) is shown as an effective approach. However, most of the existing methods improve the recognition efficiency by increasing the feature extraction dimension, ignoring the dynamics of EMG and making interpretable recognition difficult. In this paper, we propose a framework for FoG recognition based on deterministic learning. Firstly, signal is filtered out by empirical modal decomposition (EMD) in frequency bands that are not related to FoG. Next, the dynamic features of the binary phase trajectories consisting of the effective EMG of the two legs are extracted by deterministic learning. Finally, FoG was detected by random forest. Validation on an EMG dataset containing 15 patients with PD revealed that the model proposed in this paper has the best classification accuracy of 77.78%. In visualising the binary phase trajectories it was found that the FoG were cluttered while the normal were regular. Therefore, it is verified that the deterministic learning effectively improves the recognition accuracy in FoG detection.