Voxel Space Fall Detection
|
The problem of activity recognition is addressed using features acquired from a 3D reconstruction of the observed object in voxel space. A rule based recognition system is presented, that can support elderly and handicapped people in their independent living. Home environments that are able to automatically monitor the state of health of their residents can drastically improve quality of life and reduce health care costs. The experimental system implements a fall detector based on properties of a persons voxel reconstruction.
The starting point is extracting persons silhouettes in the input images. This is basically a classification problem, classifying each pixel in background and foreground. Therefore a motion detection algorithm, with shadow and highlight removal was applied.
Calibrated cameras are used to compute the 3D Voxel representation of the silhouettes. The intersection of the projection cones of the silhouettes in 3D space defines the volume of the person. A voxel based reconstruction procedure has been chosen: As a preprocessing step, the list of voxels that project in a pixel is computed for each pixel in the camera views. For the actual reconstruction itself it is then sufficient to check the silhouette images and toggle the voxels that correspond to the silhouette .
Figure 1 outlines the general reconstruction procedure. Since the voxel representation is in world coordinates, the size of the reconstructed object corresponds to it’s real world size.
In the current system two states are identified: normal and fall. The fall state is
characterized by the voxel person having a small change in position and a low height,
while having a high width or length. This simple rule based approach can easily be
extended to a sophisticated linguistic activity summarization system capable of
handling multiple states.