P 3 - 3 Pedestrian Identification and Traffic Analysis Report

Our algorithm consists of modeling and subtracting the background in the region of interest, then detecting people in every frame. Detected people are then associated to previously seen people in order to track them. Our algorithm is summarized in 11 steps and described in more detail in the comments submitted with P3-2. We now briefly describe some key steps and compare them to other methods in the literature. Model the Background. We keep a notion of recent pixel variance and update the background model when the pixels have stabilized. We use recursive averaging to update the model smoothly. (This is the technique used in Lankton P1). A very common approach for background modeling is the mixture of Gaussians approach presented by Stauffer and Grimson [1] where each pixel is modeled by several Gaussians simultaneously. This helps accommodate repetitive motion and changing scenes. Our method also operates on the pixel-level, but our method only keeps one background model. Thus it may not handle repetitive motion as well. This is seen in the " library " sequence where the slight motions of the stationary people 1