Improving the quality of university sports assessment by integrating training behavior modeling with posture sequences and spatiotemporal attention mechanisms

Overview This repository provides a reference implementation of the Dynamic Posture Behavior Integration Network (DPBIN) for university sports assessment. The framework jointly models training behavior and posture sequences and enhances evaluation through spatiotemporal attention mechanisms. It draws inspiration from recent research on integrating posture sequences, behavior features, and attention-based temporal reasoning to improve sports performance assessment in higher education. The goal is to offer a compact yet extensible codebase that supports research, teaching, and prototype development for intelligent sports assessment systems. Key Ideas The Dynamic Posture Behavior Integration Network is built on three core components: Posture Sequence Encoding Pose keypoints are extracted frame by frame and converted into feature sequences representing joint coordinates, velocities, and basic kinematic information. Training Behavior Modeling Clip-level behavior statistics—such as speed, acceleration variance, rhythm consistency, and repetition counts—are incorporated as auxiliary features. Spatiotemporal Attention Integration A recurrent backbone processes posture features over time, while an attention module highlights movement phases and behavior patterns that are most informative for quality evaluation. Together, these components generate a continuous quality score and, optionally, a discrete quality grade for each sports clip. Datasets The original study employs two university sports datasets that contain both video and posture information. Dataset Samples Classes Avg. Length (s) Annotations UATB (University Athlete Training Behavior) 7,200 6 6.0 Activity label, quality score, expert remarks SPSD (Sports Posture Sequence Dataset) 9,600 3 5.5 Pose sequence, quality level Method Summary The implemented model follows the conceptual workflow described in the original research: Posture Encoder A lightweight projection network converts per-frame pose data into latent posture embeddings. Behavior Encoder A linear transformation maps behavior statistics into a behavior embedding. Temporal Backbone A bidirectional LSTM models temporal dynamics and captures contextual information across frames. Spatiotemporal Attention Attention weights are computed using both posture and behavior embeddings, producing context vectors that emphasize important frames and behavior attributes. Fusion and Prediction Temporal context and behavior embeddings are fused into a clip-level representation.A regression head predicts a continuous quality score, and an optional classification head outputs discrete quality labels. Repository Layout File/Folder Description model.py Implementation of DPBIN with recurrent backbone and attention fusion dataset.py CSV-based dataset loader and tensor construction logic train.py Training driver script inference.py Prediction script for new samples utils.py Metrics, logging, and checkpoint utilities Example Metrics The following table summarizes representative results reported in the study: Method Acc (%) Prec (%) Rec (%) F1 (%) AUC MAE RMSE Dataset ResNet-50 91.4 90.7 89.5 90.1 95.1 0.73 0.89 UATB TimeSformer 92.6 91.8 91.3 91.5 96.0 0.70 0.86 UATB ST-GCN 93.1 92.0 91.8 91.9 96.5 0.68 0.84 SPSD Two-Stream AGCN 93.5 92.4 92.2 92.3 96.7 0.66 0.82 SPSD DPBIN (proposed) 94.9 93.8 93.6 93.7 97.9 0.59 0.74 Both These metrics serve as reference benchmarks. Full reproduction requires access to the original datasets and preprocessing steps. Extensibility Potential extensions include: Graph-based encoders for pose sequences Joint regression–classification multi-task learning Calibration-aware training for improved confidence estimation Integrating expert annotations into the attention computation Deploying the model for real-time assessment in online environments License This project is available under the MIT License. Citation If this repository contributes to your research or teaching, please cite the original study on training behavior modeling, posture sequence analysis, and spatiotemporal attention for university sports assessment.