Initialization

Most of the literature on the visual tracking system do not discuss much about how the initialization is done. Therefore, the rough estimate of the initial configuration of the object is obtained manually. Figure 1 shows that the initial estimate of the state is intentionally biased to show that the system performance is not sensitive to the initial estimate. From the estimate, the more accurate initial prior can be obtained using curve fitting algorithm described in [2]. Figure 2 shows that the accurate mean for the initial prior is obtained. The initial prior can be assumed to be Gaussian without loss of generality, and its covariance matrix is determined heuristically from the test data. To generate the sampled set representation of the density, N samples are generated using the standard random sampling method with their corresponding weights computed from the mean and the covariance -  .

Figure 1: Initial estimate of object contour

Figure 2: Result of curve-fitting
 

Tue Mar 14 02:05:32 PST 2000