To address the problem of single-modal signals being prone to misclassification in dynamic tasks such as forward, backward, and turning, this paper proposes a multi-stream neural network model based on the fusion of motion posture signals and surface electromyography (sEMG) signals. Lower-limb motion posture and sEMG signals are collected and processed using a gait cycle segmentation and alignment strategy. Based on the heterogeneity of data characteristics in terms of temporal structure, signal type, and expression dimensions, feature information is constructed, including dynamic temporal features of motion posture, frequency-domain octave spectrum of sEMG, and time-domain histogram statistics. A Transformer is used to capture the dynamic evolution of motion posture signals, and a two-way multilayer perceptron (MLP) is used to extract local response characteristics of sEMG in the time and frequency domains, forming a multi-stream neural network structure with complementary feature representations to achieve the fused expression of motion posture and local muscle fiber activation information. Results show that the proposed model improves recognition accuracy by 8%, 6%, and 24% in forward, backward, and turning tasks, respectively, significantly reducing misclassification caused by blurred motion boundaries and demonstrating superior accuracy and generalization performance. This method provides research reference and technical support for the application of multimodal fusion in complex gait recognition, and provides a reference for the design of high-precision gait recognition systems.