Ergebnisse
The following research has been done in the course of the project:
2D Silhouette Properties
The 2D approach accompishes activity reconition based on motion extraction and tracking. The motion segmentation of multiple overlapping views is tracked and region sizes are extracted. Falls are detected based on the halt of regions and their size properties, notably the width to height ratio of the objects bounding box. The recognition of the individual views is fused using a simple majority vote.
Voxel Space Fall Detection
Based on the output of the region of interest extraction of the 2D Silhouettes approach a 3D Reconstrction of the object is computed. Using multiple calibrated cameras with overlapping views, allows the backprojection of the silhouttes in 3D Voxel Space. On this recostruction a rule based fall detection is applied. Features for desicion making are object height, width to depth ratio and change in acceleration.
Spatio-temporal Local Features
In this research unit recognition in videos is done looking for salient events in the spatio-temporal domain. The visual data are static bird's view surveillance videos. We build interest points which are both salient in space and time and aim to label those interest points to represent different events: "Usual" behavior should be compared to domestic accidents like falling down on the floor. Interest points are described with local third order time jets. Classification is done using event signatures with a simplified "bag of word" approach with a spatio-temporal vocabulary. Parameters are learnt by the performance of the recognition rate.