Topology preserving self-organizing map of features in image space for trajectory classification

Abstract

Self-Organizing maps (SOM) are able to preserve topological information in the projecting space. Structure and learning algorithm of SOMs restrict the topological preservation in the map. Adjacent neurons share similar vector features. However, topological preservation from the input space is not always accomplished. In this paper, we propose a novel self-organizing feature map that is able to preserve the topological information about the scene in the image space. Extracted features in adjacent areas of an image are explicitly in adjacent areas of the self-organizing map preserving input topology (SOM-PINT). The SOM-PINT has been applied to represent and classify trajectories into high level of semantic understanding from video sequences. Experiments have been carried out using the Shopping Centre dataset of the CAVIAR database taken into account the global behaviour of an individual. Results confirm the input preservation topology in image space to obtain high performance classification for trajectory classification in contrast of traditional SOM.