A single frame of surveillance video. A person mid-stride, weight on the left foot, right arm forward. Freeze it there. Can you identify them? Probably not. The body shape is generic. The posture is ambiguous. Now play the next thirty frames. Suddenly the rhythm emerges. The timing of the heel strike. The arc of the swing. The cadence. That sequence is the signal. One frame is a letter. The full stride is the word. Last lecture established that deep learning shifted gait recognition from handcrafted features to learned representations. But that shift raised a new question: how does a model handle the fact that gait unfolds across time? A still image has no rhythm. A sequence does. Now, the key idea is that gait recognition depends not on body configuration at one instant but on how configurations transition from one instant to the next. That transition is where identity lives. Think of a gait cycle as a sentence with a fixed grammar. It begins at initial foot contact, runs through stance, crosses toe-off, completes swing, and returns to contact. Cadence counts the steps per minute. Stride time measures the elapsed interval between consecutive contacts of the same foot. Stance time runs from contact to toe-off. Swing time runs from toe-off back to contact. Single-support and double-support periods sit inside that structure. Spatiotemporal gait analysis organizes all of this into rhythm, phase, variability, pace, and base-of-support domains. Each domain is a dimension of the temporal signal. A standard convolutional network reads one frame at a time. It has no memory. Show it frame forty, and it has forgotten frame one. A recurrent neural network is different. It processes sequential inputs by updating an internal hidden state that carries information from earlier time steps forward. For gait, that memory matters enormously. The heel strike at frame one shapes what the ankle does at frame fifteen. A CNN-plus-LSTM pipeline handles this by splitting the labor: convolutional layers extract spatial information from individual frames, and the LSTM models temporal dependencies across them. One system, ST-DeepGait, uses a human skeletal graph to learn local spatial dynamics and a multilayer recurrent architecture to model sequential gait-cycle information. Here is a practical challenge. Continuous walking footage has no clean start or stop. A system must cut it into analyzable pieces. Gait-cycle segmentation is critical because sequences with different starting points, walking speeds, or frame counts must be aligned before temporal comparisons are reliable. Walking speed compounds the problem. Speed changes timing and movement patterns across every sensing modality. [short pause] Counterintuitively, using too long a sequence can hurt performance. Longer clips accumulate more variation in speed, clothing, and environment, which dilutes the stable identity signal rather than reinforcing it. Two findings here are worth holding onto, Jordan. Gait-cycle timing can be informative even when spatial appearance is discarded entirely. Approaches that analyze temporal and spatial information separately have demonstrated this. The rhythm alone carries identity. Second, when occlusion corrupts part of a sequence, an LSTM-based reconstruction method can fill the missing frames while imposing temporal consistency with the periodic structure of the gait cycle. The model knows what a stride should look like across time, and it uses that knowledge to repair gaps. That is a remarkable capability. Remember this: a single frame cannot identify a person. A sequence can. Recurrent networks and temporal models capture the transitions between body configurations, and those transitions are the identity signal. The key idea is that gait is a time series, not a photograph. Cadence, phase timing, stride-to-stride variability, and joint-angle trajectories all unfold across dozens of frames per cycle. Systems that model that flow, rather than averaging it away, extract the richest biometric signal the body produces. Next, Jordan, we turn to the real-world conditions that challenge every temporal model: clothing, fatigue, aging, and the environments where gait recognition actually has to work.