.png){kind=icon}
.png){kind=icon}
Return
TrajectoryNet plus plus : An intelligent attention system for real-time ball trajectory tracking
.png){kind=icon}
PRE
AI
Z
W
Z
DOI：10.1016/j.image.2026.117538.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In sports video analysis and related applications, accurate trajectory prediction of fast-moving small targets is essential for reliable judgment and intelligent analysis. However, such targets are frequently affected by motion blur, occlusion, and complex backgrounds, making stable and continuous trajectory modeling highly challenging. Most existing approaches rely on implicit deep feature learning, which often overlooks fine-grained geometric structures in small-object scenarios. Moreover, limited cross-layer feature interaction tends to weaken target saliency and causes fragmented trajectories, restricting robustness and generalization. To address these limitations, we propose TrajectoryNet++, a real-time trajectory tracking framework for small targets in dynamic scenes. Built upon TrackNetV2, the proposed method explicitly integrates structural priors and cross-layer semantic modeling to enhance small-target perception under complex backgrounds. Specifically, a Multi-scale Surround Prior Extraction (MSPE) module is introduced at early network stages to reinforce edge and contour cues across multiple directions and scales, mitigating saliency degradation caused by motion blur and background clutter. Furthermore, a Cross-Scale Contextual Attention (CSCA) module enables adaptive information interaction across different semantic levels in both spatial and channel dimensions, promoting effective collaboration between local structural details and high-level semantic representations. This design improves trajectory continuity and temporal stability. Experimental results on multiple sports video datasets show that TrajectoryNet++ consistently outperforms existing methods while maintaining a lightweight architecture and real-time efficiency, achieving up to 0.921 accuracy, 0.906 F1 score, and 46.2 FPS inference speed, demonstrating strong practical applicability. Our code is available at https://github.com/ZiruiDu/TrajectoryNet-.git.
Keywords:
Tennis trajectory prediction
Small object tracking
TrajectoryNet plus plus
MSPE
CSCA
Journal
S
IF:
2.7
Papers:
18
Citations:
0
Organization
Cited Papers