Activity Segmentation

Activity segmentation aims to automatically divide continuous streams of data, such as video or sensor readings, into meaningful segments representing distinct activities or events. Current research emphasizes unsupervised approaches, leveraging transformer networks and techniques like temporal optimal transport to learn segment boundaries without labeled data, and incorporating both frame-level and segment-level information for improved accuracy. These advancements are driving progress in applications like anomaly detection in driving videos and improving human activity recognition in healthcare, where accurate and efficient segmentation is crucial for analysis and diagnosis.