Instance Segmentation and Fine-Grained Contour Extraction of Basketball Players from Videos under Complex Occlusion Scenarios
Traitement du Signal · 30.04.2026 · 43 (2)
DOI: 10.18280/ts.430238 ↗Kurz gefasst
Fachartikel in der Zeitschrift Traitement du Signal, online erschienen am 30.04.2026. Verfasst von Delong Jia.
Abstract (Originalsprache)
In basketball game videos, challenges such as frequent limb bending, dense multi-player occlusion, and complex court lighting conditions severely degrade the performance of conventional instance segmentation methods, often resulting in coarse mask boundaries, distorted contours in occluded regions, and insufficient representation accuracy. To address these issues, this paper proposes a unified framework that integrates dynamic deformable feature extraction, parametric shape encoding, and continuous distance field optimization for precise basketball player instance segmentation and contour refinement. First, a boundary-enhanced multi-scale feature extraction network based on Dynamic Snake Convolution (DSC) is constructed to accurately capture slender and curved limb structures while enabling bidirectional fusion of semantic and edge features. Second, departing from the traditional binary discrete mask prediction paradigm, a two-stage continuous contour prediction mechanism is introduced, which combines Elliptic Fourier Descriptors for global shape encoding with a Signed Distance Function (SDF) field for local boundary refinement, achieving compact shape representation and sub-pixel-level contour restoration. Furthermore, a depth-hierarchy-aware copy-paste data augmentation strategy with spatial constraints is designed to simulate realistic occlusion scenarios, and a level-set evolution-based active contour algorithm is incorporated for post-processing refinement in heavily occluded areas. Finally, a multi-task joint weighted loss function is formulated to enable collaborative optimization across multiple branches. Extensive experiments conducted on both a self-constructed basketball occlusion dataset and publicly available sports vision benchmarks demonstrate that the proposed method significantly outperforms state-of-the-art instance segmentation approaches in terms of segmentation accuracy and contour fitting quality under severe occlusion conditions. The proposed approach effectively accomplishes accurate and smooth player instance segmentation, providing reliable visual technical support for practical applications such as intelligent tactical analysis
Zitieren
APA
Jia, D. (2026). Instance Segmentation and Fine-Grained Contour Extraction of Basketball Players from Videos under Complex Occlusion Scenarios. Traitement du Signal, 43(2). https://doi.org/10.18280/ts.430238
Vancouver
Jia D. Instance Segmentation and Fine-Grained Contour Extraction of Basketball Players from Videos under Complex Occlusion Scenarios. Traitement du Signal. 2026;43(2). doi:10.18280/ts.430238
BibTeX
@article{Jia2026,
author = {Jia, Delong},
title = {Instance Segmentation and Fine-Grained Contour Extraction of Basketball Players from Videos under Complex Occlusion Scenarios},
journal = {Traitement du Signal},
year = {2026},
volume = {43},
number = {2},
doi = {10.18280/ts.430238}
}
EUVIMED