Algorithmic Association Methods for Multi-Object Tracking

Motivation

Multi-object tracking (MOT) is a core problem in computer vision, aiming to detect and consistently identify multiple objects across video frames. MOT plays a crucial role in applications such as autonomous driving, surveillance, sports analytics, robotics, and behavioral research. While early approaches relied on algorithmic methods such as min-cut, Viterbi, or k-shortest paths algorithms for data association, these have largely been replaced by simpler online methods like SORT and DeepSORT. Recent trackers typically use frame-to-frame association via the Hungarian algorithm or deep-learning based association, achieving strong results on standard benchmarks. However, these methods often struggle with long-term occlusions and fragmented trajectories, where global reasoning could offer an advantage.

Project

This project investigates whether classical, pre-deep-learning algorithmic association methods can achieve competitive performance when combined with modern detection and re-identification (ReID) models. The goal is to assess if these methods are able to achieve competitive performance given the advances in computer vision and if framing tracking as a global optimization task is beneficial on current benchmarks. Based on the results, we may investigate combining well-performing, classical algorithms with graph neural networks (GNNs).

A MOT framework in python and all relevant features (detections, ReID-features, etc.) will be provided. Only the association methods need to be implemented/trained.

Thesis

Within the context of this project, several research questions could be explored. Depending on personal interests or your own ideas, the focus of the thesis can be shifted accordingly.

• How do classical association algorithms perform when paired with modern deep-learning-based features?
• In which scenarios do these methods excel or fail compared to contemporary heuristic trackers?
• Can hybrid approaches, e.g., incorporating graph neural networks (GNNs), further improve global association?

Prerequisites

The following skills are helpful. Deep learning knowledge is not required.

• Familiar with Python (Pytorch, Numpy, Scipy would be helpful)
• Basic linear algebra, optimization and discrete mathematics knowledge

Contact

To apply please email Jan Frederik Meier stating your interest in this project and detailing your relevant skills. A part of this project could be also a lab rotation.