Online Layered Multiple Object Tracking Using Residual-Residual Networks

When dealing with multiple object tracking in the real world, it faces several challenges: (a) The number of targets to be tracked will change over time, (b) The data association of the target at different times will be affected by occlusion, (c) The problem of estimating the continuous state of all targets and deciding whether the targets leave the screen and then stop tracking. In this paper, a novel multiple object tracking method that consists of a residual-residual network and a four-layer data association scheme. The residual-residual network combines a deep residual classification network and a deep residual feature network. The deep residual classification network is used to remove unwanted background noise from the frame and corrects the target position of the missing ones. It can accurately track the position of multiple targets, link their positions in each time period, combine layered target data association method, and stepwise pair the trajectories and target candidates according the features generated from the deep residual feature network. The experiments using MOT16 that is a multi-object tracking database, show that the proposed method leads most existing researches in several evaluation criteria including the accuracy, speed and false positive.