Application of a Two-Dimensional Regression Network Algorithm Model Based on Local Constraints in Human Motion Recognition

Lijun Wang, Zixu Wang, Lijuan ZhouIEEE Access 2024 · DOI

TL;DR

This paper introduces an improved 2D regression network for human pose estimation that moves beyond treating body joints as isolated points. By integrating local structural constraints (modeling the body as a graph) and relational networks (learning dependencies between joint features), the authors significantly improve detection accuracy, particularly for joints that are frequently occluded or difficult to track, like wrists and ankles.

What Problem Does This Paper Address?

Human Pose Estimation (HPE) is the task of locating specific joint points (elbows, knees, etc.) in images or video. While deep learning has made massive strides here, many current 2D regression models suffer from a "spatial blindness." They often treat each joint as an independent entity, ignoring the biological reality that joints are connected in a rigid, hierarchical structure.

This lack of structural awareness leads to two main issues:

  1. Inaccurate Feature Maps: Without knowing how joints relate, the model might produce "impossible" poses.
  2. Occlusion Sensitivity: When a limb is hidden behind an object or another body part, standard models often fail because they lack the contextual "logic" to infer a joint's position based on its neighbors.

Methods

The authors propose a multi-layered approach to inject "body logic" into a standard regression architecture:

  • The Backbone: The model uses a modified architecture featuring an Inception v4 stem for initial feature extraction, followed by a series of Hourglass-style modules (Block1 and Block2). The authors found that a depth of 8 modules provided the optimal balance between model complexity and accuracy.
  • Local Constraints (The Graph Model): Instead of just predicting coordinates, the researchers modeled the human body as an undirected graph $H = (B, F)$, where vertices are joints and edges represent physical connections. They introduced a specific penalty term in the loss function that simulates this structure, forcing the network to learn the spatial distribution and physical limits between connected nodes.
  • Relational Networks: To handle occlusion, the authors embedded a relational module that extracts appearance and geometric features from all target joints. This allows the model to use "global context"—for example, using the position of the shoulder and elbow to "guess" the location of a hidden wrist.
  • Hybrid Loss Function: The training process utilizes a combination of L1/L2 loss for coordinate accuracy, cross-entropy for probability distribution, and a structural loss term to enforce the graph constraints.

Datasets

The model was trained and evaluated using the MPII (Measurement Point II) Human Pose Dataset, a standard benchmark in the field containing around 25,000 images with over 40,000 annotated people. The primary evaluation metric used was PCKh (Percentage of Correct Keypoints - Head), which considers a prediction correct if it falls within a certain fraction of the head size.

Key Insights and Ideas

  • Structure as a Regularizer: By adding a graph-based penalty term, the authors essentially turned biological reality into a mathematical constraint. This prevents the model from making "un-human" predictions.
  • The "Sweet Spot" for Depth: The study found that accuracy peaks at 8 regression modules (88.7% PCKh). Beyond this, the gains diminish, suggesting a limit to how much "stacking" helps without better architectural innovations.
  • Context Overcomes Occlusion: The most significant "aha" moment is the effectiveness of the relational network. By allowing the feature map of one joint to "talk" to others, the model can maintain high accuracy even when a joint is not visually present in the frame.

Key Results

The researchers compared their improved model (Algorithm B) against a baseline regression model (Algorithm A) and other state-of-the-art methods:

  • Overall Accuracy: Algorithm B achieved a [email protected] score of 92.5% on the MPII dataset, outperforming several existing models and trailing the current state-of-the-art by only 0.2%.
  • Joint-Specific Improvements: The model showed high precision across the board:
  • Head: 98.8%
    • Shoulder: 95.3%
    • Wrist: 89.4% (a traditionally difficult joint)
    • Knee: 90.1%
  • Occlusion Performance: In practical tests involving indoor and outdoor environments, the improved algorithm remained consistent with manual labels, whereas comparison algorithms showed significant errors when limbs were obstructed.

Limitations and Future Work

While the model performs admirably, the authors acknowledge that large-scale occlusion remains a challenge. If a significant portion of the body is hidden, the relational network eventually runs out of "clues" to reconstruct the pose accurately.

Future research directions suggested include:

  • Improving the estimation logic specifically for "massive occlusion" zones.
  • Expanding the model for use in Virtual Reality (VR) and Augmented Reality (AR), where real-time, occluded-joint tracking is critical for immersion.
  • Refining the integration of 3D pose enhancement to further bridge the gap between 2D images and 3D motion recognition.

Summary updated

Abstract

As the behavior analysis of human body is more and more used in the fields of intelligent monitoring and motion analysis, it is of great significance to conduct research.The current two-dimensional regression network algorithm models in human motion recognition and estimation do not consider the interrelationships between human joint points, resulting in missing connections between joint points and low accuracy of feature maps. Therefore, this study proposes an improved two-dimensional regression network algorithm model based on local constraints and relational networks, and verifies its effectiveness. The experimental results show that, considering only local constraints, the proportion of the head in the correct key points of the improved algorithm in the wrist joint score is 84.72%, while the comparison algorithm is 84.55%, an increase of 1.17%. The maximum value is 88.7% when the number of regression network modules is 8. In practical applications, the actual label results of indoor and outdoor environments are basically consistent with those of the detected image, but there are errors under indoor occlusion conditions. Considering both local constraints and relational networks, the improved algorithm has variant standard scores of 98.8%, 95.3%, 93.3%, 89.4%, 95.1%, 96.2%, and 94.2% for the correct percentage of 7 joint points, respectively, which are higher than the comparison algorithm. Overall, the proposed two-dimensional regression network algorithm based on local constraints and relationship networks has practicality and effectiveness, which can be effectively applied in practical human motion recognition.
BibTeX
@inproceedings{wang2024application,
  author = {Lijun Wang and Zixu Wang and Lijuan Zhou},
  title = {{Application of a Two-Dimensional Regression Network Algorithm Model Based on Local Constraints in Human Motion Recognition}},
  booktitle = {IEEE Access},
  year = {2024},
  doi = {10.1109/access.2024.3368869},
}
Figure