MELANOSOME TRACKING USING PREDICTION BY SUPPORT VECTOR REGRESSION AND REVISION BY APPEARANCE FEATURES

摘要

Elucidations of transit port of matters in cells are veryrnimportant for finding the cause of disease. However,rntracking and detecting particles in cells are still donernmanually. Thus, we propose a melanosome tracking methodrnwhich predicts position using Support Vector Regressionrn(SVR) and revises the position using appearance features.rnAt first, we predict position of melanosome by SVR which isrntrained using Haar-like features around melanosome at timernt-1 and t. However, prediction itself is not perfect, and wernrevise the position predicted by SVR. We detectrnmelanosomes by SVM and compare intensity betweenrnmelanosome at time t-1 and neighboring melanosomes ofrnthe predicted position at time t. We select the melanosomernwith minimum intensity difference. To evaluate the accuracy,rnwe used 31 normal melanosomes and 13 melanosomes ofrnGriscelli syndrome. Our method using only SVR achievesrn89.1% for the task in which the position in time t isrnpredicted from time t-1. When we revise the position usingrnintensity features, the accuracy is improved to 97.8%. Whenrnwe give correct position at only the first frame in test videosrnand the position in all remaining frames is predicted, therntracking accuracy is 87.3% which is much higher thanrnconventional method.