MoSA-Det: motion state adaptive object detection for sports videos

Area of Science:

• Computer Vision
• Machine Learning
• Sports Analytics

Background:

• Object detection in sports videos is crucial for broadcasting and analysis.
• Existing methods struggle with motion blur and large object displacements.
• These challenges degrade feature distinctiveness and temporal aggregation accuracy.

Purpose of the Study:

• To propose MoSA-Det, a framework that uses motion states to adapt detection strategies.
• To jointly optimize feature extraction and temporal fusion for improved sports video object detection.
• To enhance accuracy in scenarios with fast-moving objects and motion blur.

Main Methods:

• Developed a Motion-Aware Adaptive Feature Module (MAAF) using motion state estimation and dynamic convolutions.
• Incorporated deformable convolutions in MAAF to handle motion blur in high-speed regions.
• Designed a State-Guided Temporal Aggregation Module (SGTA) for selective feature fusion based on motion states.

Main Results:

• MoSA-Det demonstrated improved performance on SoccerNet-Tracking and SportsMOT datasets.
• Achieved notable gains in mAP@0.5 and mAP@0.75 compared to strong baselines.
• Validated the effectiveness of the motion state adaptive strategy in sports video object detection.

Conclusions:

• MoSA-Det successfully addresses key challenges in sports video object detection.
• The proposed motion state adaptive strategy enhances both feature extraction and temporal fusion.
• This framework offers a significant advancement for applications requiring accurate athlete tracking and tactical analysis.