Efficient Multi Object Tracking Using RNN and Siamese Re-Identification

This work presents a comprehensive study and methodology to cope the problem of multi-object tracking. The methodology proposed here aims to simplify the problem space by using RNN as a regression network for predicting current object tracker information from past data and using Siamese network for re-identification. The number of re-identification is reduced by doing suboptimal perfect matching using a tree data structure for geometric location data of objects in view port.