Experiments of Several Shadow Removal Algorithms Sample Clauses

Experiments of Several Shadow Removal Algorithms. To design a high performance shadow removal algorithm, several methods were investigated to perform shadow removal in real time. The first method involved a dual- pass application of the Otsu automatic thresholding method (Otsu, 1979) to the intensities of the detected foreground pixels. The first application of the Otsu method separated the pixels into high and low intensity populations. The high intensity population was considered to be the vehicle of interest. However, the low intensity population might have consisted of both shadow pixels cast by the vehicle as well as self-shadow areas on the vehicle that were hidden from direct illumination sources. To separate these two areas, a second thresholding of the lower intensity population was performed. Those pixels above the resultant threshold were considered to be self-shadow pixels, while those pixels occupying the absolute lowest pixel range were considered to be the cast shadow. One of the primary advantages of this method is that it is computationally inexpensive. Figure 5-3 illustrates that although the method performed well for bright vehicles with dark cast shadows, it did not perform as well when darker vehicles were considered. Notice that in the latter case, self shadow regions of the pickup truck were classified as shadow and subsequently replaced by pixels from the background. Although this problem could be mitigated by allowing only points connected to the exterior of the vehicle to be classified as shadow, the algorithm also did not perform well when the cast shadow was not uniform and not as dark. This is illustrated in Figure 5-4, which was taken when the altitude of the sun was low, resulting in shadows that may have occupied higher intensity ranges than the vehicle itself.