A camera watches a parking garage entrance. Low light. The subject wears a heavy coat. The silhouette system struggles. Confidence: forty percent. Then a wrist-worn sensor adds its signal. Confidence jumps to eighty-seven. Same person. Same walk. Two sensors reading the same body from different angles. That jump is what multi-modal fusion actually does. Last lecture established that emotional states alter gait parameters in measurable ways — speed, arm swing, smoothness — but those same features also carry identity, fatigue, and pain. Now the key idea is this: no single sensor captures all of that cleanly. Multimodal gait recognition combines complementary representations — silhouettes, skeletal pose, depth, optical flow, infrared imagery, or inertial measurements — rather than relying on one signal alone. Think of fusion as a decision made at four possible moments in the pipeline. Data-level fusion combines raw sensor streams before any processing. Feature-level fusion concatenates modality-specific representations before classification, letting a model learn cross-modal relationships — but it requires compatible timing, scaling, and dimensionality. Score-level fusion combines recognition scores from separate systems, preserving specialized extractors and staying practical when sensors are heterogeneous. Decision-level fusion combines final outputs — votes or identity hypotheses — and can stay operational even when sensors use entirely separate pipelines. For example, video silhouettes describe body-movement outline. Depth maps add three-dimensional structure. IMUs provide acceleration and angular-motion signals that remain available when a camera view is obstructed or poorly lit. A published study fused video with three-dimensional inertial data specifically to address recognition under different walking conditions. That means when the coat kills the silhouette, the IMU keeps reading. The sensors cover each other's blind spots. A practical multimodal pipeline must synchronize sensors. A temporal offset between an inertial waveform and the corresponding video movement creates false cross-modal relationships — the system learns noise, not identity. And missing sensors are a central deployment problem. UGaitNet was designed to handle multiple or missing inputs — grayscale, optical flow, depth, and silhouettes — instead of assuming every modality is always present. [short pause] That design philosophy is called graceful degradation. The system keeps working when one channel goes dark. The numbers make the case. Fusing RGB silhouettes with depth or LiDAR-projected depth produced Rank-1 accuracy improvements of 19.3 and 19.7 percentage points, respectively, under a challenging clothing condition. UGaitNet reported almost perfect recognition on selected same-view or elapsed-time benchmark protocols using CASIA-B and TUM-GAID — though those results are protocol-specific and should not be read as universal real-world accuracy. Here is the counterintuitive part, Jordan. Adding more modalities can reduce trustworthiness rather than improve it. Combining an apparently low-information signal with another modality may create a stronger identifying representation than either signal alone. That means fusion increases the amount and sensitivity of personal data collected — and makes privacy risk harder to assess. A consent model appropriate for one sensor may be wholly inadequate when that sensor is fused with two others collected at different times and locations. The practical lesson is this: combining gait with complementary signals can improve recognition reliability when their errors are partly independent. But resilience and restraint must travel together. The takeaway is that multi-modal fusion is elegant engineering — and a serious ethical responsibility.