The Parkinson’s disease gait classification method proposed in this paper consists of five steps. Firstly, gait data collection includes gait videos of both normal individuals and Parkinson’s patients provided by the Gait Laboratory of Fujian University of Traditional Chinese Medicine. After gait data collection, preprocessing of the data is conducted, including data augmentation, cropping, resizing, and adding dynamic blur. Next, the Attention-LSTM model is constructed to effectively capture long-term dependencies in time series. After training and testing, the model can achieve effective classification of normal individuals and Parkinson’s patients based on gait video data, yielding results. The experiments demonstrate that the model constructed in this paper outperforms baseline models and previous Parkinson’s disease classification studies based on video data. Our research reduces the difficulty and cost of gait data collection, enabling contactless gait analysis. This will further reduce the complexity and cost of Parkinson’s disease diagnosis and lay a solid foundation for remote diagnosis of Parkinson’s disease.