Fast Multi-Object Tracking-by-Detection Using Tracker Affinity Matrix

Objectives: This study is an attempt to develop a fast real-time multi-object tracking algorithm with competitive precision and accuracy in crowded scenes. Methods/Statistical Analysis: This research extends multi-object tracking-by-detection framework by utilizing physical characteristics of trackers and their affinity for tracker guidance, in addition to detections and representation models. Guidance is done by particle filter which weights are updated through reworked observation model, featuring a term based upon tracker affinity. Tracker affinity is also used for tracker grouping that removes redundant trackers following the same target, and triggering a special propagation mode during occlusions that prevent identity switches. Findings: Our research has shown that trackers affinity matrix and algorithm features based on it yield substantial extra accuracy for tracking-by-detection framework, while taking a minor fraction of processing time. Improvements/Applications: In this paper we have introduced an approach that makes it possible to use less accurate, but faster detectors and representation model classifiers, enabling real-time processing, while keeping competitive precision and accuracy.