On-Line Kernel-Based Tracking in Joint Feature-Spatial Spaces

We will demonstrate an object tracking algorithm that uses a novel simple symmetric similarity function between spatially-smoothed kernel-density estimates of the model and target distributions. The similarity measure is based on the expectation of the density estimates over the model or target images. The density is estimated using radial-basis kernel functions which measure the affinity between points and provide a better outlier rejection property. The meanshift algorithm is used to track objects by iteratively maximizing this similarity function. To alleviate the quadratic complexity of the density estimation, we employ Gaussian kernels and the fast Gauss transform to reduce the computations to linear order. This leads to a very efficient and robust nonparametric tracking algorithm. More details can be found in [2]. The system processes online video stream on a P4 1.4GHz and achieves 30 frames per second using an ordinary webcam. 1 Similarity Between Distributions This demonstration presents a real-time object tracking system running on a PC with an ordinary webcam. The tracking of objects in a video stream is a common task, in which a model image is translated, rotated, and (possibly) deformed to match the given target images. It is important for many computer vision applications such as humancomputer interaction, surveillance, smart rooms and medical imaging. Our approach is based on the matching in the feature spaces (we used RGB color space) which are described by the probability density functions (pdf ). The pdf is estimated in the feature spaces using kernel density estimation (see Figure 1):