Instance Segmentation and Fine-Grained Contour Extraction of Basketball Players from Videos under Complex Occlusion Scenarios

With the rapid iteration of computer vision technologies [1] and their deep integration with the smart sports industry, intelligent sports event analysis has become a core driving force for the digital transformation of the sports industry [2, 3]. As one of the most globally popular competitive sports, basketball game analysis is gradually upgrading from traditional manual statistics to an intelligent and refined direction, and the core supporting role of precise visual perception of athletes is becoming increasingly prominent. However, the dynamic and complex environment of basketball courts brings many intractable technical challenges to athlete instance segmentation and contour extraction [4, 5]: athletes' limbs exhibit slender and curved non-rigid shapes, and actions such as torso twisting and limb stretching during movement lead to constant changes in target morphology; in high-intensity confrontation scenarios, densely arranged players are prone to large-scale limb cross-occlusion, with some areas even experiencing complete occlusion, resulting in incomplete target features; fluctuations in court lighting intensity, motion artifacts, and blurring effects caused by fast motion further exacerbate image quality fluctuations; meanwhile, there are significant scale differences between athletes, referees, and coaches on the court, coupled with rapid displacement, further increasing the difficulty of feature extraction and contour localization. As the underlying core technology of intelligent basketball game analysis, the performance of instance segmentation and precise contour extraction directly determines the accuracy and reliability of upper-level applications such as player trajectory tracking, quantitative action scoring, tactical decomposition, and intelligent event commentary. Therefore, conducting research on athlete instance segmentation and contour refinement for complex occlusion scenarios in basketball game videos holds significant practical necessity [6-8]. From a theoretical perspective, this study can improve the theoretical system of continuous contour representation for non-rigid human targets under complex dynamic occlusion scenarios, break through the accuracy bottleneck of traditional discrete mask segmentation, and establish a new visual segmentation paradigm integrating deformable convolution feature extraction and parametric shape priors [9, 10], enriching the research content and technical paths in the field of sports-specific visual image processing. From an application perspective, the research results can be directly implemented in professional basketball game intelligent analysis systems [11], providing coaching teams with accurate tactical optimization bases; they can be applied to campus basketball teaching correction [12], assisting students in standardizing movement postures; they can also be extended to scenarios such as public fitness posture assessment and sports big data intelligent analysis [13, 14], possessing extremely high engineering implementation value and industrial promotion potential.

Currently, scholars at home and abroad have conducted extensive research on instance segmentation, sports human segmentation, contour extraction, and occlusion optimization, forming a series of research results. However, specific research targeting complex occlusion scenarios in basketball game videos still suffers from many common defects, making it difficult to meet actual application demands. Mainstream instance segmentation models rely on pixel-level binary mask output; their inherent discrete nature leads to obvious stepped aliasing on segmentation edges, making sub-pixel level contour extraction impossible [15, 16]. In areas where athletes' limbs cross and occlude, pixel misclassification is prone to occur, leading to contour boundary distortion. Traditional deformable convolutions adopt an independent sampling point offset mechanism lacking continuous path constraints, making them unsuitable for adapting to the complex geometric shapes of basketball players' limb bending and torso twisting. Their integrity in extracting long-distance limb features is insufficient, leading to feature fracture problems. Existing contour extraction methods mostly focus on local edge detection [17, 18], lacking global human shape prior constraints. In areas where occlusion leads to missing contours, contour deformation and limb fracture are highly likely to occur, making it impossible to achieve accurate restoration of complete contours. General image data augmentation strategies do not consider the spatial depth logic of the court and cannot simulate the real physical occlusion relationship of "near occluding far." Synthetic training samples differ greatly from actual court scenes, resulting in insufficient generalization capability of models in complex occlusion scenarios. Furthermore, existing segmentation frameworks lack dedicated loss constraints for sports human targets [19]; the optimization weight distribution for edge features and shape features is unreasonable, and the collaborative fitting effect of multi-task branches is poor, making it difficult to balance segmentation accuracy and contour refinement [20]. The existence of these problems severely restricts the application and promotion of athlete instance segmentation and contour extraction technologies in intelligent basketball game analysis, necessitating the proposal of a targeted solution.

Aiming at the above research deficiencies, this paper conducts research on athlete instance segmentation and contour refinement for complex occlusion scenarios in basketball game videos. The core innovations and contributions are as follows: designing a boundary-enhanced multi-scale Dynamic Snake Convolution (DSC) feature extraction backbone network, introducing cumulative offset constraints to match athlete limb features, and realizing deep fusion of semantic and edge features; proposing a dual-stage continuous contour prediction framework combining Elliptic Fourier global shape encoding and Signed Distance Function (SDF) local refinement to achieve compact contour representation and completion in occluded areas; constructing a position-aware depth-constrained copy-paste data augmentation method to improve model occlusion robustness; designing a multi-constraint fusion active contour post-processing mechanism to optimize contour refinement; building a multi-task joint weighted loss function to achieve collaborative optimization of all branch tasks.

The organizational structure of the subsequent chapters of this paper is as follows: Chapter 2 elaborates in detail on the athlete instance segmentation and contour refinement method proposed in this paper, including the specific design of the feature extraction network, the dual-stage contour prediction mechanism, the occlusion-aware training strategy, and the multi-task joint optimization loss function; Chapter 3 verifies the effectiveness and superiority of the proposed method through multiple sets of comparative experiments and ablation experiments, and provides a detailed analysis and discussion of the experimental results; Chapter 4 summarizes the full text's research work, objectively analyzes the limitations of the research, and looks forward to future research directions.

3.1 Overall experimental setup

The experiment takes athlete instance segmentation and contour refinement under complex occlusion scenarios in basketball game videos as the core objective, verifying the effectiveness and superiority of the proposed method through joint datasets. The dataset consists of a self-built basketball game complex occlusion dataset and public sports human vision datasets. The self-built dataset contains 5000 basketball game images, covering different lighting intensities, occlusion degrees, and athlete motion postures, with each image finely annotated for athlete instances, segmentation masks, and SDF ground truth. The public datasets selected are COCO-Sports and Sports-1M, filtering relevant samples of basketball events and supplementing contour annotations to ensure data diversity and annotation consistency. The dataset is divided into training, validation, and test sets at a ratio of 8:1:1. During the training process, online data augmentation is used to generate synthetic occlusion samples to improve the model's generalization ability.

Three core evaluation metrics are adopted to comprehensively quantify model performance: instance segmentation metrics include Average Precision at IoU threshold 0.5 (AP50), Average Precision at IoU threshold 0.75 (AP75), and mean Intersection over Union (mIoU), measuring segmentation accuracy and global pixel matching degree under different IoU thresholds, respectively; contour refinement metrics include Average Boundary Distance Error and Sub-pixel Contour Fitting Error, quantifying the accuracy and smoothness of contour extraction; model efficiency metrics adopt Single-image Inference Frames Per Second (FPS) to evaluate algorithm real-time performance. The software and hardware environment is uniformly configured: hardware uses a single NVIDIA A100 GPU with 32GB memory; software is based on the PyTorch 1.12 deep learning framework, equipped with Python 3.8, OpenCV 4.5, and CUDA 11.6, ensuring the reproducibility of experimental results.

3.2 Core ablation experiments

To verify the effectiveness of each core module of the proposed method, 5 groups of ablation experiments are designed. The settings and quantitative results of each group are as follows. All data are average results on the test set, where the occlusion degree is divided into mild (occlusion ratio ≤ 30%), moderate (30% < occlusion ratio ≤ 60%), and severe (occlusion ratio > 60%).

3.2.1 Feature extraction backbone network ablation experiment

Experiments compare the performance of different feature extraction backbones to verify the role of DSC and the Semantic-Edge Fusion module. The experimental results are shown in Table 1.

It can be seen from Table 1 that as the occlusion degree increases, the segmentation accuracy of various backbone networks shows a downward trend, and contour extraction errors gradually increase, while the proposed complete feature backbone performs best under all occlusion degrees. Compared with the native ResNet-50, the proposed backbone improves AP50 by 9.5 percentage points, mIoU by 10.1 percentage points, and reduces the average boundary distance error by 1.12 pixels in severe occlusion scenarios, indicating that DSC can effectively capture the bending limb features of athletes and solve the problem of traditional convolution sampling fracture. Compared with the traditional deformable convolution backbone, the proposed backbone further improves edge feature extraction accuracy through semantic-edge bidirectional fusion, with AP75 increasing by 5.6 percentage points and sub-pixel contour fitting error decreasing by 0.23 pixels in moderate occlusion scenarios. The performance of the multi-scale network without an edge branch is slightly better than that of the traditional deformable convolution backbone but lower than that of the proposed complete backbone, verifying the important role of the edge perception branch in boundary detail optimization. Although the FPS of the proposed backbone decreases slightly due to the introduction of the DSC and edge fusion modules, it remains above 25 FPS, meeting real-time inference requirements.

Table 1. Ablation experiment results of feature extraction backbone networks

3.2.2 Dual-stage contour representation framework ablation experiment

Experiments compare the performance of different contour representation methods to verify the necessity of the dual-stage framework combining Elliptic Fourier global encoding and SDF local refinement. The experimental results are shown in Table 2.

Table 2. Ablation experiment results of the dual-stage contour representation framework

Data from Table 2 show that the proposed dual-stage continuous contour prediction framework significantly outperforms the other three contour representation methods in all indicators, especially in severe occlusion scenarios. Due to its discrete nature, traditional binary mask prediction yields rough contour edges, the highest contour distortion rate reaching 12.5% under severe occlusion, and larger boundary errors. Single SDF field prediction can achieve continuous contour extraction but lacks global shape constraints, making it prone to contour offset in occluded areas, with a contour distortion rate still reaching 9.6%. Elliptic Fourier shape prediction focuses on the global topological structure, but the local detail fitting accuracy is insufficient, resulting in boundary errors and distortion rates higher than those of single SDF field prediction. The proposed dual-stage framework combines global shape encoding and local distance field refinement. In severe occlusion scenarios, compared with traditional binary mask prediction, the average boundary distance error is reduced by 0.91 pixels, the sub-pixel contour fitting error is reduced by 0.64 pixels, and the contour distortion rate is reduced by 4.7 percentage points. This verifies that the dual-stage continuous representation mode can effectively solve the problems of insufficient discrete mask accuracy and poor robustness of single continuous representation, achieving collaborative optimization of global structure and local details.

3.2.3 Comparative experiment of occlusion-aware data augmentation strategies

Experiments compare the impact of different data augmentation strategies on model performance to verify the effectiveness of the position-aware depth-constrained copy-paste augmentation method. The experimental results are shown in Table 3.

Table 3. Comparative experimental results of occlusion-aware data augmentation strategies

It can be seen from Table 3 that various data augmentation strategies can improve model performance, among which the proposed depth-hierarchy occlusion augmentation strategy achieves the best results. Without any data augmentation strategy, the model performs worst in severe occlusion scenarios, with AP50 being only 72.1%, indicating that insufficient occlusion scenarios in training samples will lead to weak model generalization ability. Conventional geometric transformation augmentation can only improve the model's adaptability to posture and scale changes, with limited optimization effects on occlusion scenarios; although random copy-paste augmentation can increase the number of occlusion samples, it lacks depth constraints, causing synthetic samples to not conform to real occlusion logic, resulting in only a 3.5 percentage point increase in AP50 under severe occlusion scenarios. The proposed depth-hierarchy occlusion augmentation generates synthetic samples that conform to the physical law of "near occluding far" through depth estimation and geometric correction. In severe occlusion scenarios, compared with no data augmentation, AP50 increases by 5.7 percentage points, mIoU increases by 5.8 percentage points, and the average boundary distance error decreases by 0.58 pixels, significantly improving the segmentation accuracy and contour extraction capability of the model in dense mutual occlusion scenarios, verifying the rationality and effectiveness of this augmentation strategy.

3.2.4 Effectiveness verification experiment of active contour post-processing module

Two sets of controlled experiments with and without post-processing were set up to quantify the optimization effect of the level set evolution active contour post-processing module on contour refinement extraction. The experimental results are shown in Table 4.

Table 4. Effectiveness verification experimental results of the active contour post-processing module

Data from Table 4 show that after introducing the active contour post-processing module, contour refinement indicators are significantly optimized, while inference efficiency remains within a reasonable range. Without post-processing, predicted contours are prone to aliasing and local concavity-convexity in severely occluded areas, with contour smoothness being only 0.25 and the average boundary distance error reaching 3.05 pixels. After introducing the post-processing module, through energy function constraints and level set evolution, contour smoothness increases by 44%, and the average boundary distance error and sub-pixel contour fitting error decrease by 0.38 pixels and 0.23 pixels, respectively, with the optimization effect being more obvious in severe occlusion scenarios. Although the post-processing module increases inference time by 6.3 ms per frame and the FPS drops to 22.5 FPS, it still meets the real-time analysis requirements for basketball game videos. Moreover, through block-wise evolution optimization, the increase in time consumption has been effectively controlled, achieving a balance between accuracy and efficiency.

To verify the instance boundary recovery capability of the proposed method when facing complex occlusion, non-rigid posture changes, and lighting disturbances in basketball game videos, this paper conducts a visual analysis of the step-by-step processing results in typical scenarios. As can be seen from the results in Figure 5, the original input contains interference factors such as large-area overlapping of player bodies, slender limb bending caused by layup movements, and edge weakening caused by strong light shadows. Traditional segmentation results based on local texture or discrete masks are prone to adhesion, boundary fracture, and local contour distortion at occlusion junctions. In contrast, the proposed method enhances continuous responses along the bending direction of limbs through DSC in the feature extraction stage, forming more stable boundary high-response regions after semantic-edge fusion. In the dual-stage continuous contour prediction stage, global shape encoding provides reasonable human topological constraints for occluded parts, and local distance field regression further ensures the continuity and smoothness of boundary transitions. After active contour post-processing, the aliasing, protrusions, and unnatural depressions in the initial zero level set are significantly corrected. The final segmentation mask and bright green fine contour can still remain clearly separated in multi-player close-range areas. Detailed magnified results further show that contours relying solely on SDF are prone to inward or outward concavity in occluded areas, while after introducing Fourier shape priors, the contour can be smoothly completed along the natural structure of the human body.

Figure 5. Visual comparison chart of processing effects at each stage of the proposed method

3.2.5 Horizontal comparison experiment with mainstream advanced algorithms

Current mainstream instance segmentation and contour extraction algorithms were selected for comparison. Comprehensive performance comparisons were completed on a unified test set to verify the comprehensive superiority of the proposed method. The experimental results are shown in Table 5.

Table 5. Horizontal comparison experimental results of mainstream advanced algorithms

It can be seen from Table 5 that the proposed method achieves the best overall performance among all comparison algorithms, especially showing significant advantages in severe occlusion scenarios. As a traditional instance segmentation algorithm, Mask Region-based Convolutional Neural Network (R-CNN) relies on binary mask output, resulting in low contour extraction accuracy, with AP50 being only 70.1% under severe occlusion and large boundary errors. The Sparse Instance Activation for Real-Time Instance Segmentation (SparseInst) algorithm has the fastest inference speed but lacks sufficient segmentation accuracy and contour refinement in severe occlusion scenarios. Although the Human Contour Transformer algorithm focuses on contour extraction, it lacks optimization designs for occlusion scenarios, resulting in a sub-pixel fitting error of 1.31 pixels under severe occlusion. The generic SDF segmentation algorithm achieves continuous contour extraction but lacks dedicated shape prior constraints for basketball players, leading to limited performance improvement. Through the collaborative optimization of core modules, the proposed method achieves AP50 of 89.7%, 84.5%, and 77.8% in mild, moderate, and severe occlusion scenarios, respectively, which are 5.0, 7.0, and 7.7 percentage points higher than Mask R-CNN. Under severe occlusion, the average boundary distance error is reduced by 0.48 pixels and the sub-pixel contour fitting error is reduced by 0.34 pixels compared to the generic SDF segmentation algorithm. Although the FPS of the proposed method is slightly lower than that of Mask R-CNN and SparseInst, it still meets real-time inference requirements and demonstrates the best comprehensive performance in terms of accuracy and efficiency, verifying the applicability and superiority of the proposed method in complex basketball occlusion scenarios.

3.3 Model inference efficiency and engineering practicality analysis

To verify the engineering application value of the proposed method, a disassembly analysis of the inference time consumption of each functional module of the model was conducted. The results are shown in Figure 6.

Figure 6. Disassembly of inference time consumption for each module of the model

It can be seen from Figure 6 that the feature extraction backbone network and the dual-stage contour prediction network are the main sources of inference time consumption, accounting for 42.0% and 34.4% of the total time consumption, respectively. The active contour post-processing module accounts for 14.2% of the time consumption, and its time consumption has been controlled within a reasonable range through block-wise evolution optimization; other modules account for only 9.4% of the time consumption, having a relatively small impact on the overall inference efficiency. The overall single-frame inference time is 44.5 ms, and the FPS reaches 22.5 FPS, which can meet the real-time intelligent analysis requirements for basketball game videos (usually 25 FPS).

Analysis of the model lightweight transformation space indicates that the feature extraction backbone network can further compress parameters through pruning, quantization, and other methods; the DSC module can adopt a lightweight separable convolution structure; the dual-stage contour prediction network can optimize the number of Transformer decoder layers and attention heads. It is estimated that the total inference time consumption can be reduced by more than 20%, and the FPS can be increased to more than 28 FPS, further enhancing engineering practicality.

In summary, under the premise of ensuring high-precision segmentation and contour refinement extraction, the proposed method possesses good real-time inference performance and has considerable room for lightweight transformation. It can meet the actual engineering application requirements of professional basketball game intelligent analysis and sports posture assessment, and has high industrial promotion value.