This invention discloses a method,
system, and medium for human
fall detection based on multi-view geometry. The method includes: acquiring intrinsic and extrinsic parameters of multiple fixed cameras and a ground
homography matrix; simultaneously acquiring images and obtaining a binary
mask of the
human body with global identifiers; determining radial lines based on the
nadir points of each camera, adaptively selecting the optimal viewing angle camera, extracting foot and head image points, and calculating the horizontal position and vertical height using the
homography matrix and geometric relationships; using multi-view ground
homography constraints, mapping
mask pixels in a reference view to other views for consistency
verification, backprojecting verified pixels onto the
ground plane to reconstruct the contact area and calculate the area; and fusing at least the temporal changes in height and contact area to determine whether a fall has occurred. This invention can accurately measure key physical quantities related to falls and has advantages such as strong robustness, high accuracy, good
interpretability, and controllable hardware costs.