Summary of the Presentation

Enhancing Multiple Object Tracking Accuracy via Quantum Annealing
(Published in Scientific Reports, 2025 – NEC Solution Innovators, Ltd., Yasuyuki Ihara)

This presentation introduces a practical study applying quantum annealing to improve the speed and accuracy of multiple object tracking (MOT) — a fundamental task in image recognition used in areas such as traffic monitoring, autonomous driving, and manufacturing automation. Traditional MOT algorithms face a persistent trade-off between accuracy and speed: classical optimization or deep learning methods can achieve one but rarely both simultaneously. The study demonstrates that quantum annealing, together with ensemble and reverse-annealing strategies, can overcome this limitation.

The presentation begins with a review of the Ising model, which provides a physical analogy for combinatorial optimization problems. In this model, each variable (or “spin”) takes a value of +1 or –1, and the lowest-energy configuration represents the optimal solution. Quantum annealing leverages quantum tunneling, which allows the system to escape local minima and reach the global optimum more efficiently than classical optimization.

The MOT problem is formulated as bipartite graph matching between consecutive video frames, ensuring a one-to-one correspondence between detected objects. This formulation can be expressed as a Quadratic Unconstrained Binary Optimization (QUBO) problem, making it suitable for quantum annealing hardware such as D-Wave Advantage2. The study’s objectives are threefold:

1. Accelerate MOT using quantum annealing.
2. Improve accuracy through ensemble integration of multiple trackers.
3. Enhance efficiency using reverse annealing, which refines the previous frame’s solution for the next frame.

The proposed Quantum-Ensemble MOT combines outputs from several trackers into a unified QUBO graph, allowing quantum annealing to integrate multiple hypotheses without a large increase in computation time. Compared with majority-voting integration, a new cyclic integration method continuously reduces error rates as more trackers are added, improving robustness against detection noise and occlusion.

Experiments used the UA-DETRAC dataset for vehicle tracking, with YOLOv5 as the detector and DeepSORT as the baseline method. Performance was evaluated using several metrics:

 MOTA (overall tracking accuracy),
 IDF1 (identity consistency),
 IDSW (number of identity switches),
 APE (counting accuracy), and
 TTS (Time to Solution) for computational efficiency.

Results on test videos demonstrated that the QA-based ensemble achieved perfect counting (APE = 0) and fewer identity switches, particularly in scenes with many visually similar vehicles or severe occlusion. The introduction of Reverse Annealing (RA) further accelerated computation. Because consecutive frames are similar, RA reuses the previous solution as an initial state and refines it within microseconds. Compared to conventional quantum annealing, RA achieved over 99 % reduction in computation time while maintaining or improving accuracy, making real-time MOT feasible.

The combination of QA-Ensemble and RA proved capable of simultaneously achieving high accuracy and low latency, demonstrating the potential of quantum annealing for real-world smart systems. Potential applications include traffic signal optimization, collision prediction in autonomous vehicles, and defect detection in manufacturing.

Remaining challenges include hybrid integration with classical computing, robustness under adverse conditions (e.g., night or rain), and the development of larger-scale quantum hardware. Nevertheless, this work provides evidence that quantum computing can move beyond theory to practical deployment in transportation and automation, forming a foundation for future smart-city systems.