One-Stage Object Detection and Feature Embedding Network for Multiple Object Tracking

In real environments, it is very important not only to detect objects in images but also to track their movements robustly. Object detection and data correlation are essential to track multiple objects. With advances of deep learning, rapid performance improvements have been achieved in the object detection field over the past decade, and this has significantly contributed to multi-object tracking accuracy. On the other hand, deep leaning-based feature embedding has been researched in the data association for the past several years. Many previous studies have applied multiple object tracking by performing two different tasks independently or through multiple stages. In this paper, we propose a one-stage object detection and feature embedding network. The unified network integrates a feature embedding sub-network into a one-stage object detection network. We train the detection network using a supervised learning method and the feature embedding network using a self-supervised learning method through multi-task learning. Our experimental results show that the proposed multiple object tracking framework using the unified network gives both better accuracy and faster speed.