Related Experiment Videos
Bio-Inspired Enhanced Adaptive Centered Collision Optimizer for Hyperparameter Optimization of Multi-Scale
1Sports Department, Beijing Institute of Technology, Beijing 100081, China.
Biomimetics (Basel, Switzerland)
|July 27, 2026
Summary
This study introduces a new biomimetic optimization framework for accurate boxing action recognition. The Enhanced Adaptive Centered Collision Optimizer improves deep learning models, achieving superior performance in combat sports analysis.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Biomimetic Computing
Background:
- Accurate boxing action recognition is crucial for intelligent combat training and injury prevention.
- Existing deep learning methods struggle with high-speed motions, background interference, and manual hyperparameter tuning.
- The original Centered Collision Optimizer has limitations in population diversity and adaptive regulation for complex optimization tasks.
Purpose of the Study:
- To develop a novel biomimetic optimization-driven framework for enhanced boxing action recognition.
- To address limitations in feature extraction, robustness, and hyperparameter optimization in current deep learning models.
- To improve the accuracy and stability of boxing action recognition systems.
Main Methods:
- Proposed a Multi-Scale Spatio-Temporal Adaptive ConvNeXt (MSTA-ConvNeXt) network integrating multi-scale dynamic deformable convolution, Bi-GRU, and dual-channel attention.
- Developed an Enhanced Adaptive Centered Collision Optimizer (EACCO) with Tent chaotic elite opposition-based initialization, adaptive convergence factor, and hybrid mutation.
- Utilized EACCO to automatically optimize MSTA-ConvNeXt hyperparameters for improved performance.
Main Results:
- Achieved high accuracy (96.1%) and F1-scores (95.8%) on the Boxing Jab Skeleton Dataset.
- Attained excellent results (95.4% accuracy, 95.0% F1-score) on the Olympic Boxing dataset, surpassing state-of-the-art methods.
- Ablation studies confirmed the effectiveness of EACCO components and its superiority over manual tuning and other optimizers.
Conclusions:
- The proposed biomimetic optimization framework significantly enhances boxing action recognition accuracy and robustness.
- EACCO provides an effective solution for hyperparameter optimization in complex deep learning models for sports analysis.
- This research offers a promising direction for intelligent sports action recognition systems.