Simple Online and Realtime Tracking (1602.00763v2)

Abstract: This paper explores a pragmatic approach to multiple object tracking where the main focus is to associate objects efficiently for online and realtime applications. To this end, detection quality is identified as a key factor influencing tracking performance, where changing the detector can improve tracking by up to 18.9%. Despite only using a rudimentary combination of familiar techniques such as the Kalman Filter and Hungarian algorithm for the tracking components, this approach achieves an accuracy comparable to state-of-the-art online trackers. Furthermore, due to the simplicity of our tracking method, the tracker updates at a rate of 260 Hz which is over 20x faster than other state-of-the-art trackers.

• The paper introduces a simple tracking framework leveraging basic techniques like Kalman filters and the Hungarian algorithm to prioritize detection quality.
• It integrates CNN-based detection, particularly Faster R-CNN variants, achieving a notable 33.4% MOTA at 260 Hz for real-time tracking.
• The open-sourced, minimalistic approach establishes a high-performance baseline for advancing pedestrian tracking and autonomous systems.

Overview of "Simple Online and Realtime Tracking" (SORT) Paper

The paper "Simple Online and Real-time Tracking" (SORT) by Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft presents a straightforward yet effective approach to the problem of multiple object tracking (MOT). The motivation behind the paper is to create a tracking framework that is not only accurate but also efficient enough for real-time applications such as pedestrian tracking in autonomous vehicles.

Key Components and Design Philosophy

The approach championed in the paper is characterized by its simplicity and efficiency. The authors abandon complex appearance models and instead rely on fundamental techniques such as the Kalman Filter for motion prediction and the Hungarian algorithm for data association. They emphasize the significance of detection quality in tracking performance, suggesting that improving the object detector can lead to substantial performance gains.

Primary Contributions

The contributions of this paper are multi-faceted:

• Leveraging CNN-based Detection: The authors integrate state-of-the-art convolutional neural network (CNN) based detection frameworks, specifically Faster R-CNN (FrRCNN), to enhance tracking performance.
• Minimalistic Tracking Framework: The proposed approach employs a basic yet effective tracking-by-detection methodology, which focuses on the bounding box coordinates for motion estimation and data association.
• Public Availability: The code is open-sourced, providing a baseline for further research and fostering wider adoption in various applications requiring real-time tracking.

Methodology

Detection

The paper evaluates the impact of detection on tracking quality by comparing different frameworks within the Faster R-CNN ecosystem. Tracking performance was tested using:

• FrRCNN(ZF): A fundamental architecture by Zeiler and Fergus.
• FrRCNN(VGG16): A deeper architecture by Simonyan and Zisserman.

Both models were tested against the traditional Aggregate Channel Filter (ACF) detector, showing a significant improvement in accuracy with the CNN-based detectors, particularly FrRCNN(VGG16).

Motion Estimation and Data Association

The methodology relies on a linear constant velocity model to predict the target’s future state, and the Kalman filter is used to handle the state updates. For associating detections with existing targets, the paper adopts the intersection-over-union (IOU) metric for computing the assignment cost matrix, resolving it with the Hungarian algorithm.

Track Management

The framework initializes new tracks for unmatched detections and terminates tracks not associated with any detection for a specified duration. The probationary period ensures robustness against false positives, while the minimalistic approach aids in maintaining real-time capabilities.

Experimental Results

The paper's experiments yielded compelling outcomes:

• The method achieved a high accuracy (MOTA) of 33.4% and impressive speed (260 Hz), outperforming many complex trackers.
• The tracking framework demonstrated robustness across various sequences, showing minimal lost tracks (ML metric).

Benchmark Comparisons

When compared with state-of-the-art online and offline methods, SORT showed competitive performance, particularly excelling in efficiency. The simplicity and speed of SORT make it suitable for real-time applications, a crucial feature for autonomous systems.

Implications and Future Work

The findings from this research underscore the importance of detection quality in tracking systems. By demonstrating that classical methods can achieve state-of-the-art results when paired with strong detection frameworks, the authors highlight a potential direction for future work. Specifically, the integration of detection and tracking into a unified framework could further enhance performance. Additionally, future studies might explore object re-identification techniques to handle long-term occlusions and reappearances, extending the practical applications of this work.

Conclusion

The SORT methodology exemplifies the principle of simplicity yielding effective results. By focusing on robust frame-to-frame associations, it emphasizes the role of detection accuracy over complex tracking models. This work not only establishes a high-performance baseline for MOT but also sets the stage for future innovations in real-time tracking frameworks. The availability of the code for public use further amplifies its impact on the research community and practical applications.