In the large-scale video surveillance systems relevant information extraction and representation processes play an important role in the interpretation of the scenes. In particular, when the amount information grows up, due to a large number of monitored areas, it could be necessary to focus the attention on a part of total available information only. In this cases, one of the main problems in event detections is to reconstruct the scene from limited observations. In this paper an innovative way of sparse information representation, based on information theory, is presented. The Self Organizing Maps (SOMs) have been employed at two different steps: for classifying and correlating observed sparse data time series. By means of Information Bottleneck it is possible to determine the best data representation (in the SOM-space) as trade-off between the capabilities to recover the signals and maintain the statistical similarities of original data. The experiments shown how the so called information bottleneck based SOM selection, for knowledge modelling, can be applied to the field of crowd monitoring for people density map estimation and event detection. The results on synthetic and also on real video sequences are presented.
INFORMATION BOTTLENECK-BASED RELEVANT KNOWLEDGE REPRESENTATION IN LARGE-SCALE VIDEO SURVEILLANCE SYSTEMS
CHIAPPINO, SIMONE;MARCENARO, LUCIO;REGAZZONI, CARLO
2014-01-01
Abstract
In the large-scale video surveillance systems relevant information extraction and representation processes play an important role in the interpretation of the scenes. In particular, when the amount information grows up, due to a large number of monitored areas, it could be necessary to focus the attention on a part of total available information only. In this cases, one of the main problems in event detections is to reconstruct the scene from limited observations. In this paper an innovative way of sparse information representation, based on information theory, is presented. The Self Organizing Maps (SOMs) have been employed at two different steps: for classifying and correlating observed sparse data time series. By means of Information Bottleneck it is possible to determine the best data representation (in the SOM-space) as trade-off between the capabilities to recover the signals and maintain the statistical similarities of original data. The experiments shown how the so called information bottleneck based SOM selection, for knowledge modelling, can be applied to the field of crowd monitoring for people density map estimation and event detection. The results on synthetic and also on real video sequences are presented.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
