Core Generator of Hypotheses for Real-Time Flame Detecting

摘要

The flame is a visually unstable and constantly changeable process, which causes considerable difficulties for its detection in the video streams. Although the modern architecture of convolutional neural networks can show high accuracy, their integration into real-time systems is problematic, because they require a large amount of computing resources. To reduce the number of these resources, it is proposed to select possible regions of interest (ROI), which are based on the developed generator of hypotheses. Compared to existing flame detection algorithms, the developed generator of hypotheses allows you to work with the minimum of computing resources and has a high degree of classification completeness due to improved methods of color segmentation and moving objects detection.