Object detection in motion management scenarios based on deep learning.

Area of Science:

• Computer Vision
• Sports Science
• Machine Learning

Background:

• Accurate identification of athletes, equipment, and field boundaries is crucial for sports performance analysis.
• Existing object detection methods struggle with temporal information loss, multi-targeting, target overlap, and task coupling in sports scenarios.
• These limitations hinder the effectiveness of current network models for motion management detection.

Purpose of the Study:

• To develop a novel supervised object detection method specifically for motion management scenarios in sports.
• To address the limitations of existing methods by enhancing temporal information capture and feature extraction.
• To improve the accuracy and efficiency of detecting multiple targets in dynamic sports environments.

Main Methods:

• Designed a Temporal-Spatial Module (TSM) integrating temporal offset and spatial convolution to capture motion scene dynamics.
• Implemented a deformable attention mechanism to boost feature extraction for individual target actions.
• Introduced a decoupling structure to separate regression and classification tasks for improved detection performance.

Main Results:

• The proposed method achieved a mean Average Precision (mAP) of 92.298% on open-source datasets, outperforming seven other common object detection networks.
• Ablation studies confirmed that each proposed module contributes significantly to the overall detection accuracy.
• Experimental results demonstrate the method's effectiveness and superiority in motion management detection scenarios.

Conclusions:

• The novel supervised object detection method significantly enhances the accuracy of target detection in motion management.
• The TSM module, deformable attention, and decoupling structure are key innovations driving improved performance.
• This approach offers a promising solution for advanced sports performance analysis and motion management.