Smoke segmentation in video sequences

Smoke detection in outdoor scenes using video sequences is particularly important for early warning systems because smoke usually rises before flames arise. Dynamic texture features of smoke are color, shape, motion, transparency, texture. The paper presents a smoke-detection method for outdoor spaces early fire-alarming based on video processing using color, shape and motion features. The proposed approach includes two stages. Firstly, local smoke regions are detected based on motion estimation and chromatic analysis. The clustering of such local regions provides global smoke regions in a scene. At this stage, smoke and non-smoke regions are analyzed in order to exclude errors of false rejection. The suspicious region is extracted by using block-matching algorithm. Secondly, global regions are verified by using statistical and temporal features. In this research, smoke colored blocks and turbulence characteristics. For experimental researches the database of dynamic textures Dyntex and database of Bilkent University were used. Dense smoke, transparent smoke, and non-smoke videos have been used for testing the proposed method. The developed method of smoke detection on video provides 97.8-99 % of accuracy for smoke sequences. Smoke was detected without false alarms in three burns. The most remarkable aspect about the results is the algorithm’s ability to filter motion other than smoke. In fact, it can be seen from the image sequences extracted that two potential sources of false alarms like the movement of tree leaves due to wind and the movement of people crossing the scene are mostly filtered. The alarms are therefore undoubtedly triggered by the smoke arising from the burns. Smoke video image was performed to verify in experiments, the results have proved the validity of the method proposed in this paper.