Gait recognition in real-life scenes is complicated by occlusion and illumination changes. These factors can lead to recognition difficulties due to significant differences in the appearance of the target. Most existing methods rely on single-modal features and have limited depth in feature extraction. They often ignore the correlation between different feature types, which leads to performance degradation under changes in viewing angle, occlusion, or complex backgrounds. However, silhouette and skeletal features are closely dependent—for example, upper-body bone movements directly affect the body contour. To address these issues, this paper proposes MSFusion-Gait, an innovative gait recognition method. It adaptively fuses contour and skeletal features to fully exploit their complementarity. This approach maintains recognition performance even when some features are missing or occluded. Additionally, we designed three different attention mechanisms that automatically identify key information and improve recognition accuracy. In particular, the proposed spatiotemporal-attention HPP mechanism integrates information from neighboring frames and focuses on the periphery of occluded regions. This compensates for missing features and captures the spatiotemporal dynamic dependence of contour and skeletal cues. Experiments on the CASIA-B, OUMVLP-Pose, and GREW datasets demonstrate that the proposed model achieves higher accuracy and robustness, proving its effectiveness in challenging gait recognition tasks under complex environmental conditions.