A pose-induced multimodal gait recognition method was proposed and a new multimodal dataset named CASIA-V containing real 3D human poses was constructed, in order to achieve multi-view fine-grained gait recognition in complex environments. The dataset was collected using multiple synchronized devices, capturing 93 subjects walking under three conditions (normal, carrying a bag, and wearing a coat) from 11 different viewpoints. It contained 11 253 RGB sequences along with corresponding motion-capture poses and depth maps. At the methodological level, a spatiotemporal alignment strategy was introduced to match motion-capture poses with video-based poses. A separately trainable 3D Pose-Induced Module (3D-SIM) was designed to extract real pose semantics, and a dual-branch model named SIGait was constructed to fuse pose and silhouette features. The experiment showed that, under the cross-view evaluation protocol of the CASIA-V dataset (excluding samples from the same view), SIGait achieved a Rank-1 accuracy of 97.7% in the normal walking scenario. Compared with the baseline method MSAFF, SIGait improved the Rank-1 accuracy by 5.1 and 7.1 percentage points in the carrying-bag and wearing-coat scenarios, respectively. Notably, the inference process did not require motion-capture data, validating that real 3D pose effectively enhanced gait recognition in complex scenarios. The results further showed that real 3D pose semantics enhanced multimodal feature representation and maintained stable recognition performance under view variations and appearance interference.