Human motion detection and segmentation from moving image sequences

Abstract

In this paper, a method for the detection and segmentation of human motion in moving image sequences is presented. For detecting motion, the intensity of each pixel is convolved with the second derivative of temporal Gaussian smoothing function or temporal Laplacian of Gaussian (LoG) filter. The zero-crossing in a single frame of the resulting function indicates the positions of moving edges. Intensity changes in time due to small illumination effects do not produce zero crossing; thus, they are not interpreted as human motion by this method. The optical flow velocity is computed by using the spatial and temporal derivatives of this function, and it is normal to the zero crossing contours. Pixels belongings to the normal velocities are back-projected in the original color image sequences to achieve a segmented color image. Experiments show that the moving object is detected correctly and achieves a good segmentation results.