*E. Weights Computation Based on Bounding Box*

In this section shows that the knowledge of some prior information about the signal ft to be recovered enables to improve the quality of the reconstructed signal for the same number of random projections. Such prior information is introduced by means of a vector of weights w = [w(l) …w( -W)J , where w(i) denotes the weighting factor associated with the ith coefficient of the vector 6 = ФT ft in the 2-Dwavelet domain. Since 6j is not available, it propose to set the weights needed for the recovering of the foreground image at time t based on the estimated bounding boxes of the previous frame.

To be more precise, a window capturing the likely foreground position in the current frame is derived from the predicted bounding box positions and sizes. A window p is constructed on the basis of the pixel domain representation of the foreground image ft and next transformed to another window defined in the wavelet domain. The ith window coefficient p (i) is set according to the position and size of the object bounding box. For pixel locations within the bounding box, it set p (i) = 1 . For pixel locations outside the bounding box, the corresponding p (i) smoothly decays to zero as they get far from the bounding box. It compute the distances dx (i) and dy (i) to the nearest vertical and horizontal bounding box border, respectively. Then the window coefficient associated to this pixel is given by

/ \ —aw(dx(i)+dy(i)) _

p(i) = e where aw> 0 a parameter is related to the rate of decay of the window.