Deep ensemble model with improved score level fusion for suspicious action detection
Abstract
Video surveillance (VS) systems are crucial for maintaining interior and outdoor security in todays unstable environment. VS components including behavior identification, comprehension, and categorizing behaviors as normal or suspicious can be used in real-time applications. Individuals are at risk in case of harms occurred by the suspicious actions. As criminal activity rises, detection of illegal behavior is essential to reducing such instances. In the beginning days of VS, it was manually done by people and was exceedingly exhausting, since suspicious behaviors were uncommon compared to regular activities. Thereby, the proposed IoT based system work using Deep Ensemble Model (DEM) for SAD focuses to recognize the suspicious activities in videos. Converting videos to frames is performed initially at pre-processing phase. Improved BIRCH segmentation is performed after pre-processing and then extract features that include Improved Local Texton XoR Pattern (ILTXP), BOVW and MOBSIFT features. The recognition of suspicious actions is done with DEM including Bi- LSTM, DBN and DMO frameworks. The ultimate outcomes on suspicious activities are detected by improved SLF process.