Action Recognition
Action recognition is the task of identifying what activity is happening in a video or image sequence. Given a clip of someone running, cooking, or waving, the model classifies the action into a predefined category. This differs from object detection (which asks "what is here?") by focusing on temporal patterns and motion ("what is happening?").
Early approaches used hand-crafted features like optical flow histograms and skeleton joint trajectories. Modern methods rely on deep learning: two-stream networks process RGB frames and optical flow separately, 3D CNNs (C3D, I3D, SlowFast) apply convolutions across both space and time, and video transformers (TimeSformer, VideoMAE) use self-attention to capture long-range temporal dependencies. SlowFast networks are particularly popular because they process video at two frame rates simultaneously, capturing both fast motion and slow context.
Action recognition is used in surveillance (detecting fights or falls), sports analytics (classifying plays and tracking player performance), manufacturing (verifying assembly steps), healthcare (monitoring patient mobility), and human-computer interaction (gesture-based controls). Real-time action recognition on edge devices requires lightweight architectures like MoViNet or X3D, which balance accuracy with inference speed.
