Video Monitoring Queries

Recent advances in video processing utilizing deep learning primitives achieved breakthroughs fundamental problems analysis such as frame classification and object detection enabling an array of new applications. In this paper we study the problem interactive declarative query on streams. particular introduce a set approximate filters to speed up queries that involve objects specific type (e.g., cars, trucks, etc.) frames with associated spatial relationships among them car left truck). The resulting are able assess quickly if predicates true proceed further or otherwise not consider avoiding costly operations. We propose two classes $IC$ notation="LaTeX">$OD$ , adapt principles from image detection. utilize extensible neural architectures easy deploy utilize. addition, statistical techniques process aggregate involving constraints streams demonstrate experimentally increased accuracy estimation. Finally, framework based extreme value theory detect unexpected its utility. Combined these constitute robust monitoring techniques. application proposed conjunction can dramatically increase rate by at least orders magnitude. present results thorough experimental benchmark data sets scale demonstrating performance benefits practical relevance our proposals.