Video understanding is the broad field of extracting meaningful information from video sequences, going beyond single-frame image analysis to incorporate temporal dynamics, motion patterns, and event progression over time. While image recognition answers "what is in this frame," video understanding answers "what is happening, how is it changing, and what might happen next."

Core tasks within video understanding include action recognition (classifying activities like running, cooking, or assembling a part), temporal action detection (finding when specific actions start and end in untrimmed video), video object tracking (following specific objects across frames), video captioning (generating natural language descriptions of events), and video question answering (responding to queries about video content). Architectures range from 3D CNNs like SlowFast and X3D that process short clips, to video transformers like TimeSformer and VideoMAE that apply self-attention across both spatial and temporal dimensions.

Practical applications cover surveillance and security (anomaly detection, person re-identification), sports analytics (play recognition, player tracking), manufacturing quality control (detecting process deviations in assembly line footage), autonomous driving (predicting pedestrian intent, understanding traffic flow), and content moderation. Recent advances in video-language models allow natural language interaction with video content, enabling search, summarization, and question answering over large video archives without manual annotation.