Unsupervised dense crowd detection by multiscale texture analysis

Fagette Antoine (1) , Nicolas Courty (2) , Daniel Racoceanu (3) , Jean-Yves Dufour (4)

1 Thalès Solution Asia
2 OBELIX - Environment observation with complex imagery
3 IPAL - Image & Pervasive Access Lab
4 Thales Services

This study introduces a totally unsupervised method for the detection and location of dense crowds in images without context-awareness. With the perspective of setting up fully autonomous video-surveillance systems, automatic detection and location of crowds is a crucial step that is going to point which areas of the image have to be analyzed. After retrieving multiscale texture-related feature vectors from the image, a binary classification is conducted to determine which parts of the image belong to the crowd and which to the background. The algorithm presented can be operated on images without any prior knowledge of any kind and is totally unsupervised.

Mots clés: Dense crowd; Segmentation; Feature extraction; Quadtree; Diffusion maps; Multiscale

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]

Fagette Antoine, Nicolas Courty, Daniel Racoceanu, Jean-Yves Dufour. Unsupervised dense crowd detection by multiscale texture analysis. Pattern Recognition Letters, 2013, pp.1-27. ⟨hal-00904210⟩