SIMPLE ONLINE AND REALTIME TRACKING

This research delves into an efficient methodology for online and realtime object tracking, emphasizing the crucial role of detection quality in enhancing tracking performance. By employing a straightforward combination of established techniques like the Kalman Filter and Hungarian algorithm, the study achieves tracking accuracy on par with advanced online trackers. Notably, the simplicity of the proposed tracking approach enables updates at an impressive rate of 260 Hz, surpassing the speed of existing state-of-the-art trackers significantly.

Introduction

This paper introduces a streamlined approach to multiple object tracking (MOT) focusing on real-time applications by detecting and associating objects efficiently. Key points from the section include:

• MOT is approached as a data association problem wherein detections across frames need to be associated, with emphasis on efficiency for real-time tracking like in pedestrian tracking for autonomous vehicles.
• The method aims to simplify tracking processes, updating at a rate of 260 Hz, significantly faster than other state-of-the-art trackers.
• Observations from a visual MOT benchmark motivated the employed methods, highlighting MHT and JPDA techniques, as well as the significance of detection quality in influencing tracking performance.
• The study ignores appearance features beyond bounding boxes for simplicity and prioritizes efficiency over robustness to detection errors.
• Traditional data association techniques are favored over complex components to handle edge cases, with a focus on reliability in common frame-to-frame associations.
• By leveraging CNN-based detection alongside Kalman filter and Hungarian method, this minimalistic tracking approach showcases efficiency and reliability for online tracking, particularly demonstrated in pedestrian tracking scenarios.
• The paper's contributions include employing CNN-based detection in MOT, presenting a practical tracking approach based on Kalman filter and Hungarian algorithm, and providing open-sourced code for research experimentation, especially in collision avoidance applications.

Literature Review

The traditional methods for Multiple Object Tracking (MOT) like Multiple Hypothesis Tracking (MHT) and Joint Probabilistic Data Association (JPDA) filters face challenges in making decisions quickly in high uncertainty scenarios due to combinatorial complexity, making them unsuitable for realtime tracking.

• Rezatofighi et al. addressed the combinatorial complexity in visual MOT by approximating JPDA efficiently using integer programming, improving the tracking process.
• Kim et al. enhanced MHT performance by introducing appearance models for target pruning but still faced delays in decision-making, limiting suitability for online tracking.

Online tracking methods often use appearance and motion models for associating detections with tracklets, aiming to optimize one-to-one correspondences using algorithms like the Hungarian algorithm.

• Geiger et al. used the Hungarian algorithm in a two-stage process for tracklet formation and association, where geometry and appearance cues were crucial, but the approach was limited to batch computation.
• The new approach, inspired by previous methods, simplifies association to a single stage with basic cues, ensuring faster and more efficient tracking for online and realtime applications.

Methodology

The proposed method in this paper encompasses key components essential for effective multiple object tracking in online and real-time scenarios: