Auto-Adaptive Parallel SOM Architecture with a modular analysis for dynamic object segmentation in videos

This paper presents a new Background Subtraction System scheme based on two Self Organized Maps (SOM) that adapt in a parallel way at different rates. Our system can automatically identify the possible issue that mainly affects the performance of the video segmentation model (such as dynamic/static background, stationary dynamic objects, jittering camera, camouflage, etc.,) by analyzing the initial frames of the video sequence. Four different modules are implemented to treat separately all these situations and different analysis are performed on them. Our system maintains a high adaptive capability in all the video sequence analysis, it is not restricted to only the initial frames of the sequence as most segmentation algorithms. In our Auto-Adaptive Parallel SOM Architecture, AAPSA, a Suspicious Foreground analysis is constantly monitoring the segmentation results in order to obtain a reduction on the false positive rates. AAPSA was validated with Change Detection 2014 and BMC databases by using the same initial model parameters on both databases demonstrating its robustness with different and complicated scenarios. In order to simulate the videos that a security guard must analyze, 3 sequences were created by concatenated videos from AVSS, PETS2001 and Change Detection. The segmentation results obtained demonstrate that our system produced the best definition of dynamic objects compared with State of the Art models.