Multiple object tracking with head detection

Abstract

As the development and flourishing of object detection, the tracking-by-detection method has been popular and well-studied. The tracking-by-detection method has been an effective and reliable way to track either a single object or multiple objects by reading in detection sequences. And among most tracking problems, pedestrian tracking is the most popular and practical one. However, most tracking methods are based on a single detector, which results in loss of valuable image information, especially in pedestrian detection. In order to keep more information from raw images, this paper presents a simple yet relatively effective way to create an online multiple object tracking system reading from results of two detectors, namely, full-body detector and head detector. To evaluate the effectiveness of head detector, this tracking system is kept as fundamental as possible following the principle of Occam’s Razor. Therefore, a combination of two traditional yet powerful tools, Kalman filter and Hungarian algorithm, are adopted for motion model and data association, respectively. This tracking system is evaluated on detection results of a variety of scenarios and performs relatively good comparing to other tracking systems.