Police sample video coarse segmentation method and law enforcement recorder video intelligent slicing method

Abstract

The invention discloses a police sample video coarse segmentation method and a law enforcement recorder video intelligent slicing method, and relates to the technical field of video analysis. According to the method, a police sample video is preferentially divided into a plurality of confidence intervals which are relatively independent in semantic and physical contexts based on semantic coherence of a text, and when semantic information is missing or unreliable, secondary segmentation is carried out based on visual features and metadata in a seamless switching manner; and a multi-modal segmentation result is optimized based on semantic dominance and an interval merging rule, so that a relatively accurate coarse segmentation result is obtained, and the workload of manual annotation is remarkably reduced. According to the method, the intelligent slicing network fusing multi-mode and gating decision is designed, complementation is realized by utilizing multi-mode data, dynamic self-adaptive multi-mode fusion is realized by introducing the gating network, noise interference is effectively inhibited, and thus the accuracy and reliability of boundary detection of the model are remarkably improved.

Images