Multi-task Deep Learning for Fast Online Multiple Object Tracking

Abstract

We present a multi-task deep learning framework to improve the performance of the Multiple Object Tracking (MOT) problem. Motion and appearance cues are ombined together to build an online multiple object tracker. While being accurate, our tracker also runs fast enough. We have made two major contributions in this paper: (1) Learn appearance features offline with triplet loss. (2) Train a quality-aware deep network by sharing convolutional features. The proposed online tracker achieves the state-of-art performance on the UA-DETRAC dataset [17] while being efficient in terms of running speed at the same time.