Melanosome Tracking using Prediction by Support Vector Regression and Revision by Appearance Features

Mika Shimomura and Kazuhiro Hotta


Melanosome, Tracking, Detection, Support Vector Regression


Elucidations of transit port of matters in cells are very important for finding the cause of disease. However, tracking and detecting particles in cells are still done manually. Thus, we propose a melanosome tracking method which predicts position using Support Vector Regression (SVR) and revises the position using appearance features. At first, we predict position of melanosome by SVR which is trained using Haar-like features around melanosome at time t-1 and t. However, prediction itself is not perfect, and we revise the position predicted by SVR. We detect melanosomes by SVM and compare intensity between melanosome at time t-1 and neighboring melanosomes of the predicted position at time t. We select the melanosome with minimum intensity difference. To evaluate the accuracy, we used 31 normal melanosomes and 13 melanosomes of Griscelli syndrome. Our method using only SVR achieves 89.1% for the task in which the position in time t is predicted from time t-1. When we revise the position using intensity features, the accuracy is improved to 97.8%. When we give correct position at only the first frame in test videos and the position in all remaining frames is predicted, the tracking accuracy is 87.3% which is much higher than conventional method.

Important Links:

Go Back