Visual Tracking

In many video surveillance applications, cameras are fixed and we are interested in tracking the motion of the foreground, which could be people or cars. Obviously, frame difference only gives us a rough idea of which regions may contain moving objects, but such a simple method can neither sperate the foreground from the background, nor tell us which image regions are moving regions. As a result, doing simple frame differencing will not output good tracking results. Background subtraction is such a technique that the foreground region can be separated from background by maintaining a background model, and classifying each pixel into either foreground and background. The assumption here is that the camera is more or less fixed such that we can maintain or train a background model. Generally, such background model can be pixel-wise, i.e., each pixel is modeled independently. For each image pixel location, we compare the input image pixel and its background model and determine its class. A simple approach just uses a mean image as the background, those pixels which deviate from the corresponding background pixels beyond a threshold could be taken as foreground pixels. More sophisticatedly, background model can be trained by taking a sequence of the background, and difference methods can be use to represent the background model: 1. Mean: For each pixel, a mean pixel is used to represent the background, i.e., the background at time is (,). To check if an input pixel (,) is a foreground pixel or not, we just check ∣(,) − (,)∣ > or not. 2. Mean & Covariance: For each pixel, a Gaussian can be used to model the distribution of such pixel, i.e., the model for a background pixel is represented by a mean pixel 3. Mixture of Gaussian: For each pixel, instead of using a Gaussian, a mixture of Gaussian can also be used, since the pixel distribution could have multiple peaks. For example, if the background contains a computer screen, whose images will have a periodical moving strips, obviously, the distribution of the background pixel is not uni-peak. Such mixture of Gaussian model can capture such scenario very well. 4. Temporal Deviation: Since the computation of the mixture of Gaussian model is a bit intensive, the temporal deviation model has a simple representation. 2 It uses its previous frame to predict the current frame of background, this model is able to adapt to changing lighting. …