<p>This study explores the use of a Recurrent Neural Network (RNN) for classifying the three subphases of the stance phase (Initial Contact, Midstance, and Propulsion) in the Windlass mechanism during walking. The experiment was conducted using the Anthropometric Vision System (AVS), which included a 60 fps camera positioned 50 cm from the foot and interfaced with LabVIEW software. Thirty five participants, twenty men and fifteen women, were placed with blue markers on the right foot at specific anatomical points. Afterwards, the subjects were instructed to walk within a marked area. Thirty-five video frames were captured during each subphase, and the proposed RNN was trained and validated using Leave-One-Subject-Out cross-validation. The algorithm achieved validation accuracy of 0.943 with a validation loss of 0.152. Additionally, the performance metrics precision, recall and F1 score was of 0.943, 0.946 and 0.947, respectively, outperforming other tested models such as Dense Neural Networks with Long Short-Term Memory (DNN-LSTM) and Recurrent Neural Networks with LSTM (RNN-LSTM). The proposed RNN’s superior performance was attributed to its ability to learn temporal variations in gait. Furthermore, the optimized architecture supports the integration into the AVS to classify the main subphases of the stance phase and the measurement of anthropometric angles, such as the Medial Longitudinal Arch (MLA) and the Medial Joint Angle (MJA), during walking. The results demonstrate the effectiveness of the RNN for real-time classification and angle measurement in dynamic gait analysis, proving its potential for clinical applications in foot biomechanics.</p>