Using of deep convolutional neural networks for visual features extraction in multiple objects tracking task

The article explores the task of matching visual features when tracing objects on a video sequence. The authors conducted a comparative analysis of existing methods for extracting visual object features and similarity estimation for the re-identification of objects. An algorithm that compares several approaches for feature extraction and similarity estimating has been developed. The authors carried out an experimental assessment of the speed and accuracy of the algorithms using the MOT-16 and PETS 2007 datasets. It is shown that the most accurate estimates of the similarity of objects are achieved by calculating the modified value of the normalized cross-correlation function between features derived from the neural network average pooling layer.