Multiple Object Tracking Method Under the Conditions of Uncertainty of Their Detection

Abstract

In this paper, we consider the existing methods of Multiple Object Tracking (MOT), point out their advantages and disadvantages. We give a problem formulation of the MOT under the conditions of uncertainty, analyze the efficiency of different methods of prediction of the objects positions missed by the detector, such as: the method based on the Kalman filter, the moving average, the autoregression, the least squares method. We propose a structure of convolutional neural network (CNN), which allows to extract “deep features” of found and predicted visual objects on a single frame in one pass of the CNN. This structure is based on the single shot multibox detector with the addition of the ROI-pooling layer in order to extract the “deep features” of the objects. We develop a MOT method under the conditions of detection uncertainty o. It allows to achieve higher accuracy of tracking objects in comparison with the known ones and to process in real time at a speed of 24 frames per second. This method can be used for linear motion of objects and visual scenes. The accuracy evaluation was carried out using the training and test dataset 2D MOT 15.