Part-Based Visual Tracking via Online Weighted P-N Learning

Heng Fan, Jinhai Xiang, Jun Xu, Honghong LiaoThe Scientific World JOURNAL 2014 · DOI

TL;DR

This paper introduces a novel part-based visual tracking algorithm that leverages "online weighted P-N learning" to improve classifier performance by assigning weights to training samples. By segmenting objects into local feature blocks (LFBs) and tracking them individually, the method achieves robust performance, particularly in challenging scenarios like occlusion and pose changes, outperforming existing state-of-the-art trackers.

What Problem Does This Paper Address?

Object tracking, a fundamental task in computer vision with applications ranging from human-computer interaction to robotics, faces significant challenges due to factors like occlusion, illumination changes, scale variations, and pose changes. Existing tracking algorithms often struggle to maintain accuracy and robustness in these dynamic environments. Specifically, appearance models, crucial for tracking, are susceptible to these variations. While global representations can be easily affected by occlusion, local representations, though more robust, can still suffer from issues like drifting and the invalidation of feature blocks. This paper aims to address these limitations by proposing a more robust part-based tracking approach that can adapt to changing conditions.

Methods

The core of this paper's contribution lies in its novel "part-based visual tracking algorithm using online weighted P-N learning." Here's a breakdown of the key methods:

  • Part-Based Representation: Instead of tracking the entire object as a single unit, the object is first segmented into fragments. From these fragments, "local feature blocks" (LFBs) are selected to represent the object. This part-based approach is intended to be more resilient to partial occlusions and deformations.
  • Online Weighted P-N Learning: This is the central learning mechanism. It's an extension of the P-N learning framework, which learns from both positive and negative samples. The "weighted" aspect is crucial:
  • Property Weight: This weight is derived from how many times a sample has been classified as positive or negative.
  • Classification Weight: This weight is determined by the classifier's confidence in its prediction for a given sample. By assigning these weights, the algorithm aims to decrease classification errors and improve the discriminative power of the classifier for each LFB.

  • Independent LFB Tracking: Each LFB is tracked independently using its corresponding weighted P-N learning classifier, often within the framework of Lucas-Kanade optical flow.

  • Dynamic LFB Update Strategy: To handle situations where an LFB becomes invalid (e.g., due to occlusion), a "substitute strategy" is employed. When an LFB is detected as invalid, it's replaced with a new, appropriate block from a set of unused blocks, ensuring the tracking process can continue effectively.

Datasets

The paper evaluates its proposed method on thirteen challenging image sequences. These sequences are not newly introduced benchmarks but are standard datasets used in the visual tracking community to test algorithm performance under various challenging conditions. The paper does not mention specific training datasets, implying the learning is done online during the tracking process.

Key Insights and Ideas

  • The Power of Weighted Learning: The core insight is that by intelligently weighting training samples in the P-N learning framework, the classifier's ability to distinguish between target and background can be significantly enhanced. This is particularly important for online learning where the data distribution can shift.
  • Robustness through Fragmentation: Representing an object as a collection of local feature blocks, rather than a single entity, provides inherent robustness. If one part is occluded or distorted, the others can still contribute to tracking the object's overall location.
  • Adaptive Replacement for Occlusion: The dynamic replacement of invalid LFBs is a critical mechanism for handling occlusions. This prevents tracking failure when a part of the object is temporarily obscured, allowing the tracker to recover and continue.
  • Local Features for Stable Parts: The paper suggests that local feature blocks are better at representing stable parts of an object, making the classifier more discriminative and the tracking more focused.

Key Results

The paper presents both quantitative and qualitative results to demonstrate the effectiveness of their proposed method.

  • Quantitative Comparison: The proposed tracker consistently outperforms six state-of-the-art tracking algorithms (l₁ tracking, FG tracking, IVT tracking, MIL tracking, TLD tracking, and CT tracking) across thirteen challenging image sequences. This is evidenced by lower center location errors (Table 2 and Figure 8) and higher overlapping rates (Table 3). For instance, the average center location error for their method is 5.78 pixels, significantly lower than most other trackers.
  • Qualitative Comparison: The paper showcases qualitative results on various challenging scenarios:
  • Heavy Occlusion: The tracker demonstrates robustness in handling severe or long-time partial occlusions, as shown in Figure 9(a).
  • Scale Variation: The method effectively tracks objects with large scale variations and slight rotations (Figure 9(b)).
  • Fast Motion and Motion Blur: The tracker maintains performance even with fast and abrupt motion, where other trackers tend to drift (Figure 9(c)).
  • Illumination Variation: The algorithm is shown to be resilient to changes in illumination (Figure 9(d)).

Limitations and Future Work

The authors acknowledge certain limitations:

  • Nonrigid Deformation and Full Occlusion: The paper states that the algorithm's performance drops when the tracked target is nonrigid and has extremely heavy deformation, or is fully occluded for an extended period.
  • Exact Tracking in Some Scenes: While outperforming others, the algorithm still fails to track objects "exactly" in some scenes.

Future work is not explicitly detailed, but the limitations suggest potential avenues for improvement, such as developing more sophisticated methods for handling extreme nonrigid deformations and prolonged full occlusions. The current approach of replacing invalid LFBs could also be further refined.

Summary updated

Abstract

We propose a novel part-based tracking algorithm using online weighted P-N learning. An online weighted P-N learning method is implemented via considering the weight of samples during classification, which improves the performance of classifier. We apply weighted P-N learning to track a part-based target model instead of whole target. In doing so, object is segmented into fragments and parts of them are selected as local feature blocks (LFBs). Then, the weighted P-N learning is employed to train classifier for each local feature block (LFB). Each LFB is tracked through the corresponding classifier, respectively. According to the tracking results of LFBs, object can be then located. During tracking process, to solve the issues of occlusion or pose change, we use a substitute strategy to dynamically update the set of LFB, which makes our tracker robust. Experimental results demonstrate that the proposed method outperforms the state-of-the-art trackers.
BibTeX
@article{fan2014part,
  author = {Heng Fan and Jinhai Xiang and Jun Xu and Honghong Liao},
  title = {{Part-Based Visual Tracking via Online Weighted P-N Learning}},
  journal = {The Scientific World JOURNAL},
  year = {2014},
  doi = {10.1155/2014/402185},
}
Figure