To address the problems of misidentification of similar gaits, excessive feature dimensionality, and computational complexity in gait recognition, this paper proposes a gait recognition method based on feature-level fusion. After validating the complementarity between motion posture signals and surface electromyography (sEMG) signals, parameters in the time, frequency, and time-frequency domains of the two types of signals were extracted. Based on the energy distribution of acceleration, angular velocity, and angle signals from motion posture signals, feature-level fusion was performed. A dual constraint strategy combining Gain-based discriminability filtering and energy-ratio stability filtering was adopted to reduce feature dimensions, yielding the most discriminative feature subset, upon which the XGBoost model was applied for gait recognition. Experimental results showed that the proposed method improved the average recognition accuracy by 8.6% over the baseline model that used only motion posture signals, reaching 95.8%. Specifically, the accuracies for forward, backward, and turning gaits reached 89.8%, 95.2%, and 97.3%, respectively, effectively reducing the misidentification rates for these three similar gaits. Furthermore, the feature-level fusion strategy effectively improved computational efficiency, and the energy distribution-based feature selection strategy reduced the impact of background noise on feature parameter perturbations, thereby enhancing model stability. This method provides strong technical support and engineering application value for gait feature parameter identification and real-time intelligent gait recognition control of exoskeletons.