The Demo Site

For the tracking application, the estimation of the ‘true’ motion vector is crucial. The complex discrete wavelet transform (CDWT) base motion estimation algorithm produced superior results for the estimation of the dense flow field and has been evaluated. First, the comparison of the results of the Lucas and Kanade’s (LK) and Horn and Schunk’s (HS) motion estimation algorithms is performed. Second, tracking performances are compared for the cases of CDWT-based and LK-based flow field. Lastly, the tracking performance of the proposed tracker is evaluated by using a number of test sequences and is compared to the Correlation and Mean Shift Tracker. It is observed that it can successfully track various different targets and is robust to changes of the target signature.

Since CDWT is shift-variant so it cannot be used directly for motion estimation. Several modifications have been proposed to make the DWT shift-invariant. Among the methods used is Redundant Discrete Wavelet Transform (RDWT), Overcomplete Discrete Wavelet Transform (ODWT). But these 2 methods provide only invariance for integer-shifts. So Double-Density Wavelet Transform (CDDWT) is proposed. CDWT based motion estimation algorithm is robust and provides sub-pixel accuracy which is important for tracking. CDWT algorithm hierarchical structure and proceeds from coarse to fine resolution level. At each level, motion is estimated for each subpel and the resultant flow field is propagated to the next resolution level by scaling the flow vectors and wrapping the transform coefficients of the reference image accordingly. For estimating each subpel, a quantity called the subband squared difference is obtained by the sum of absolute differences of the values of the subpels in the six detailed subimages. The corresponds to a quadratic surface whose minimum gives the desired displacement. These surfaces are accumulated through the levels to obtain the ‘cumulative squared difference’. The result of the algorithm is a real-valued motion estimate for each pixel in the images. The tracking algorithm is to track any kind of target selected by operator. The target can be rigid or non-rigid and can change pose, size and shape during tracking. The optical estimator is crucial for success of tracking algorithm.
The stimulation have been performed to test different aspects of the algorithm, first, the quality and suitability of the flow field generated by the CDWT-based motion estimation algorithm have been evaluated, second, stimulations replacing the CDWT-generated flow in the tracking algorithm with the Lucas and Kanade’s flow are performed. Lastly, the proposed tracking algorithm is compared with the correlation Tracker and the Mean shift Tracker.
The suitability of the flow generated by CDWT-based motion estimation method is evaluated in three different ways. Firstly, the flow method, secondly, the proposed tracking algorithm is evaluated. From that comparison we find that CDWT based tracker is the most accurate and efficient motion tracker compared to others method. CDWT based tracker method can also maintain the track and follows the flow information successfully. Although CDWT based method not as precise as the others, but it can produced a denser and smoother flow field than other methods especially for regions where only low frequency components were present.