What makes for good multiple object trackers?

Abstract

This paper explores the importance of detection and appearance features for multiple object tracking. Extensive detectors including hand-crafted methods and deep learning methods have been tested. We found in this paper that simply improving detection performance can lead to much better multiple object tracking results. The data association methods used in this paper are Kalman Filter and Hungarian algorithm as proposed in [1]. CNN features and color histogram features are extracted as appearance features to measure similarities between objects. Our experiments show that appearance features can help with data association. We then combine detection and data association together as an overall system. The proposed system can track multiple objects at a speed of 17 fps with high accuracy.