Abstract

The value in using active contour models (snakes) as an image segmentation tool comes from the possible interactions between high level ‘intelligence’ processing models and the algorithm itself. Such a characteristic makes the algorithm a naturally attractive option in motion tracking, since when we image sequences consists of frames that are closely correlated to the frames directly preceding them. Such information from preceding frames is often not utilized by most general purpose edge detectors, hence wasting a valuable resource. With properly initialized snakes, the original location of the snake in the preceding frame should be relative close to the current location/morphology of the snake, prompting the snake to actively seek energy minimization and seek to go toward the ‘new’ location of the object and hence continuing the ‘locked’ track of the object in question.


In this project, we compared our simple implementation of William and Shah’s approach to snakes with traditional region detectors in motion tracking, such as simple thresholding as well as region growing(which lack interaction with high level models). Our results show that our simple snake algorithm performs nearly as well as traditional methods in most cases (0.817 accuracy compared to 0.840). With more tweaking of parameters, we feel confident that using snakes in motion tracking will offer vast improvement over traditional methods.