Tracking Multiple Video Targets with an Improved GM-PHD Tracker

Xiaolong Zhou, Hui Yu, Honghai Liu, Youfu Li

TL;DR

This paper introduces an improved Gaussian Mixture Probability Hypothesis Density (GM-PHD) tracker designed to tackle the challenging problem of tracking multiple moving targets in video, especially when they are in close proximity or occluded. The core innovation lies in a novel weight penalization mechanism that leverages multi-feature fusion to accurately resolve ambiguities in target association, leading to more robust and precise tracking.

What Problem Does This Paper Address?

Tracking multiple objects in video is crucial for applications ranging from robotics and surveillance to human-robot interaction. However, existing methods, including the popular GM-PHD filter, struggle when targets get close to each other. This proximity can lead to ambiguous measurements, causing trackers to assign the wrong identities to targets (mistracking) or even lose track altogether. Existing solutions often fail to handle these "close movement" scenarios effectively, particularly when multiple measurements could plausibly belong to a single target or when targets have similar appearances. This paper aims to bridge this gap by developing a more robust tracking system that can accurately distinguish and follow targets even under these challenging conditions.

Methods

The proposed tracker builds upon the GM-PHD filter but introduces several key improvements:

  • Entropy-based Birth Intensity Estimation: To combat false positives from noisy video data, an entropy-based method is used to more accurately estimate the initial appearance (birth intensity) of new targets. This helps to filter out spurious detections.
  • Weight Matrix and Ambiguous Weight Determination: A central component is the construction of a "weight matrix" that captures the association probabilities between predicted target states and incoming measurements. The paper proposes a novel method to identify "ambiguous weights" within this matrix, which arise when a single target is associated with multiple measurements (Case 1) or when a measurement is incorrectly associated with a target (Case 2). This method considers both the total weight and the predicted target states, improving upon previous approaches that only used total weight.
  • Multi-Feature Fusion for Weight Penalization: To resolve these ambiguous weights, the tracker fuses multiple features:
  • Spatial-Colour Appearance: Models the color distribution within a target's region, incorporating spatial constraints.
  • Histogram of Oriented Gradient (HOG): Captures edge and texture information.
  • Target Area: Uses the size of the target. These features are combined to create a penalty score. Ambiguous weights are then penalized based on this score, effectively reducing the likelihood of incorrect associations.

  • Mutual Occlusion Handling: For situations where targets are completely hidden from view, the paper incorporates a previously developed game-theoretical method to handle mutual occlusion.

Datasets

The paper evaluates its proposed tracker on several types of datasets:

  • Synthetic Image Sequences: Used to validate the core weight penalization method and compare it against baseline GM-PHD and CPGM-PHD trackers.
  • Outdoor Human Surveillance Scenarios: Real-world sequences to test performance in more complex environments.
  • Cell Moving Surveillance Scenarios: Datasets with high target density and close movements to assess robustness.
  • Challenging Surveillance Scenarios: Including interactive person tracking, person and luggage tracking, and crowd tracking from PETS2006 and PETS2009 datasets.

The paper does not introduce new datasets but uses existing ones to demonstrate the effectiveness of its approach.

Key Insights and Ideas

  • The "close movement" problem is critical: The paper highlights that standard GM-PHD trackers fail significantly when targets are near each other, leading to identity switches and mistracking. This is a fundamental limitation that needs addressing for practical applications.
  • Ambiguity isn't just about total weight: The authors demonstrate that simply looking at the sum of weights for a target isn't enough to identify problematic associations. Their method of considering both total weight and predicted target states is crucial for correctly identifying ambiguities in different scenarios.
  • Multi-feature fusion is key to robust association: Relying on a single feature (like color or size) is insufficient when targets are similar or occluded. Fusing spatial-color appearance, HOG, and target area provides a more comprehensive representation that allows for better discrimination and penalization of incorrect associations.
  • Handling occlusion is a separate but important challenge: While the focus is on close movement, the integration of a game-theoretical occlusion handling method shows a commitment to building a comprehensive tracking system.

Key Results

  • Improved Tracking Accuracy: The proposed tracker significantly outperforms the conventional GM-PHD and CPGM-PHD trackers, especially in scenarios with closely moving targets. This is evidenced by higher MOTA (Multi-Object Tracking Accuracy) and MOTP (Multi-Object Tracking Precision) scores.
  • Reduced Mismatch Rate: The weight penalization and multi-feature fusion approach effectively reduces the mismatch rate (MMR), indicating fewer instances of targets being assigned incorrect identities. For example, on synthetic images, the MMR dropped from 12.88% for GM-PHD to 0% for the proposed tracker.
  • Qualitative Success: Visual results show the proposed tracker maintaining correct identities for targets in close proximity and even during partial occlusions, where other methods fail.
  • Computational Cost: The tracker is implemented in MATLAB and its processing speed is shown in Figure 13. While the game-theory based occlusion handling can be computationally intensive, the overall performance is competitive for the complexity of the problem.

Limitations and Future Work

  • Computational Cost: The paper acknowledges that the processing speed is not fast enough for real-time applications, particularly due to the game theory-based occlusion handling.
  • Object Detection: The authors note that the use of a simple background subtraction method for object detection can introduce noise, which might still be tracked as targets, potentially impacting MOTA scores. They suggest that a more robust object detection method could further improve performance.
  • Future Work: The authors suggest that employing a more efficient occlusion handling method would be beneficial for real-time applications and will be explored in future work.

Summary updated

Abstract

Tracking multiple moving targets from a video plays an important role in many vision-based robotic applications. In this paper, we propose an improved Gaussian mixture probability hypothesis density (GM-PHD) tracker with weight penalization to effectively and accurately track multiple moving targets from a video. First, an entropy-based birth intensity estimation method is incorporated to eliminate the false positives caused by noisy video data. Then, a weight-penalized method with multi-feature fusion is proposed to accurately track the targets in close movement. For targets without occlusion, a weight matrix that contains all updated weights between the predicted target states and the measurements is constructed, and a simple, but effective method based on total weight and predicted target state is proposed to search the ambiguous weights in the weight matrix. The ambiguous weights are then penalized according to the fused target features that include spatial-colour appearance, histogram of oriented gradient and target area and further re-normalized to form a new weight matrix. With this new weight matrix, the tracker can correctly track the targets in close movement without occlusion. For targets with occlusion, a robust game-theoretical method is used. Finally, the experiments conducted on various video scenarios validate the effectiveness of the proposed penalization method and show the superior performance of our tracker over the state of the art.
BibTeX
@inproceedings{zhou2015tracking,
  author = {Xiaolong Zhou and Hui Yu and Honghai Liu and Youfu Li},
  title = {{Tracking Multiple Video Targets with an Improved GM-PHD Tracker}},
  booktitle = {Sensors},
  year = {2015},
  url = {https://openreview.net/forum?id=hS9gZad1z6f},
}
Figure