Abstract:To enhance the adaptability and safety of intelligent walking-assistive robots in the context of aging and low-fertility societies, and address the key issues of turning intention recognition and fall risk warning during walking, this paper proposes a gated recurrent unit prediction model for foot rotation angle based on 3D visual data. By analyzing foot rotation angle, the core parameter of lower limb movement, this model realizes accurate prediction of users’ straight and turning walking intentions. First, the 3D visual point cloud data during walking is collected, the physical model of the user’s lower limb linkage structure is constructed, and the time series of foot rotation angle feature is extracted. Subsequently, a gated recurrent unit (GRU) network is designed to perform temporal modeling of gait features to capture the dynamic change law of foot movement during walking. Finally, the walking intention is discriminated according to the threshold of foot rotation angle, and comparative verification is carried out with a variety of typical time series models, including long short term memory (LSTM), GRU_Attention, BiGRU, BiLSTM and Transformer. The proposed GRU model achieves a mean absolute error of 4.14×10-2 and a mean squared error of 5.12×10-3, and a test time of 0.04 seconds in the foot rotation angle prediction task. Compared with other models, while maintaining a smaller model size and shorter training time, the GRU model achieves lower prediction error and demonstrates good comprehensive performance. The proposed gated recurrent model based on foot rotation angle can effectively learn the dynamic gait features related to walking intention, and achieves a sound balance between prediction accuracy and computational efficiency, which provides a feasible technical basis for the subsequent integration into the real-time control system of walking-assistive robots.