Alleviate Similar Object in Visual Tracking via Online Learning Interference-Target Spatial Structure

Guokai Shi, Tingfa Xu, Jiqiang Luo, Jie Guo, Zishu Zhao2017 · DOI

TL;DR

This paper tackles the challenge of visual object tracking when similar-looking objects interfere with the target. They propose a novel approach that explicitly models the spatial relationship between the target and interfering objects, allowing for more robust tracking even during occlusion. This method achieves state-of-the-art performance on benchmark datasets.

What Problem Does This Paper Address?

Visual object tracking is crucial for many intelligent systems, but it faces significant challenges. One particularly tricky problem is "similar object interference," where an object that looks very much like the target appears in the scene. This is problematic because standard tracking algorithms often rely on appearance features and surrounding context. When a similar object is present, it can easily be mistaken for the target, leading to tracking drift or failure, especially when the target is occluded by this similar object. Existing methods struggle to differentiate between the target and these visually similar distractors, particularly when they share similar appearance and are in close proximity.

Methods

The core of this paper's contribution is the introduction of an Interference-Target Spatial Structure (ITSS) constraint into existing Correlation Filter (CF) based trackers. The proposed approach works in three main stages:

  1. Tracking based on ITSS model: This is the primary tracking phase. The system dynamically maintains a graph representing the spatial structure between the target and any detected similar objects. This graph's edges represent the spatial relationships and their thickness signifies the interference degree. This structure helps discriminate between the target and similar objects.
  2. Update Appearance Model: The appearance models for both the target and similar objects are updated collaboratively. This prevents the target's model from being corrupted by the appearance of a similar object, especially during occlusion. An "interference degree weight" is introduced to dynamically switch between a baseline tracker and the ITSS-constrained model.
  3. Re-detection of the Target: If the target is severely or totally occluded by a similar object, the ITSS constraint can become ineffective. In such cases, a re-detection strategy is employed, combining the target's and similar object's appearance models to find the lost target.

The ITSS constraint is formulated as a score function that is optimized using a structured Support Vector Machine (SVM) approach, leveraging a minimum spanning tree model for efficient computation.

Datasets

The paper evaluates its approach on two widely used benchmark datasets for visual tracking:

  • OTB-2013
  • OTB-2015

These datasets contain a variety of challenging sequences, including those with occlusion, illumination changes, deformation, and importantly, similar object interference.

Key Insights and Ideas

  • Spatial Structure is Key: The paper highlights that simply relying on appearance features is insufficient when dealing with similar objects. Explicitly modeling the spatial relationship between the target and potential distractors is crucial for robust discrimination.
  • Collaborative Model Updates: Updating the target's appearance model in isolation can lead to drift when a similar object is present. The proposed method's collaborative update strategy for both target and interference models is a significant improvement.
  • Dynamic Switching: The introduction of an "interference degree weight" allows the tracker to dynamically switch between a standard CF tracker and the more robust ITSS-constrained model, adapting to varying levels of interference.
  • Re-detection for Severe Occlusion: Recognizing that even sophisticated constraints can fail under extreme occlusion, the inclusion of a dedicated re-detection mechanism is a practical and effective solution.

Key Results

The proposed approach, termed "ITSS," demonstrates state-of-the-art performance on the OTB-2013 and OTB-2015 datasets.

  • Quantitative:
    • Achieves the best average Center Location Error (CLE) in six out of nine challenging sequences.
  • Achieves the best average Overlap Ratio (OR) in seven out of nine sequences.
  • Outperforms the baseline Staple tracker by 1.5% and 0.8% in OSR on OTB-2013 and OTB-2015 respectively, and by 2.4% and 1.3% in DP.
  • Qualitative:
    • Visualizations show that the ITSS tracker successfully maintains tracking through severe occlusions by similar objects, where other trackers often fail.
  • The re-detection mechanism is shown to be effective in recovering lost targets.
  • Speed: The algorithm runs at approximately 35 frames per second on a standard Intel i7 CPU with 4GB RAM, making it suitable for real-time applications.

Limitations and Future Work

  • Limitations:
    • The paper notes that under severe deformation and in-plane rotation occurring simultaneously with similar object appearance, the tracker might experience slight drifts.
  • When the target is almost completely occluded by a similar object, the ITSS constraint's effectiveness diminishes, necessitating the re-detection strategy.
  • Future Work:
    • The authors suggest that the ITSS constraint could be further improved by incorporating more sophisticated feature representations.
  • Extending the approach to multi-target tracking scenarios where multiple targets might interfere with each other could be a valuable direction.

Summary updated

BibTeX
@misc{shi2017alleviate,
  author = {Guokai Shi and Tingfa Xu and Jiqiang Luo and Jie Guo and Zishu Zhao},
  title = {{Alleviate Similar Object in Visual Tracking via Online Learning Interference-Target Spatial Structure}},
  year = {2017},
  doi = {10.3390/s17102382},
}
Figure