MaiSci ☕

  In this project, we are using Kalman and Particle filters to track motions of objects. (hands, pedestrians) With naive Kalman filter implementation, it is easy to track a moving pedestrian without major shape, or shadow change. Kalman filter (KF) tracking a pedestrian. Blue circle represent the distribution of the KF gaussian With particle filter, we track the face of Romney. The green frames enclose the target window, and blue points represent particles in particle filter (PF). The following frames are with different noise levels, but all get good face detections. As for tracking Romney’s left hand, things are not as trivial since the appearance of his hand keeps changing. Thus we take history into account and update the tracking window model continuously, following: Template( t ) =α Best(t) +(1-α) Template( t-1 ) Hand tracking with history updating Till now, we did not consider the change of size in tracking a target. In order to consider size, we need to let the particles in PF ha

The goal is to detect motions in a series of consecutive images. In this project, the pyramidal approach and hierarchical Lucas-Kanade algorithm are implemented. Frame interpolation is also executed with the help of the LK algorithm. We start with a simple motion towards the right. The “beginning” and “end” figures are the only two frames we have. The central block goes from the middle to the right by a certain range. We are going to extract the motions of pixels from those two figures as the green arrows show below. Two figures we have Motions extracted from the two figures From the motions we extracted, we could do frame interpolation based on the arithmetics below, and create the following dynamic gif animation. It interpolates the frames between the “beginning” and “end” figures show a consecutive motion of the central block towards the right. Then we implement another real world image interpolation. It is not perfect since the source frames (first frame and last frame shown in the