Applying TS-DBN model into sports behavior recognition with deep learning approach

• Published in: The Journal of supercomputing 2021, Vol.77 (10), p.12192-12208
• Main Authors: Guo, Yingqing, Wang, Xin
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial neural networks; Belief networks; Compilers; Computer Science; Datasets; Deep learning; Deep Learning in IoT: Emerging Trends and Applications - 2019; Interpreters; Machine learning; Markov chains; Model accuracy; Processor Architectures; Programming Languages; Sports; Video data
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-021-03772-x

The purposes are to automatically collect information about human sports behavior from massive video data and provide an explicit recognition and analysis of body movements. The analysis of multi-scale input data, the improvement of spatiotemporal Deep Belief Network (DBN), and the different pooling strategies are regarded as the focuses to improve the belief networks in deep learning (DL). Moreover, a human sports behavior recognition model is proposed based on particular spatio-temporal features. Also, video frame data are collected from the Royal Institute of Technology (KTH) and University of Central Florida (UCF) datasets for training. The TensorFlow platform is employed to simulate the built algorithm. Finally, the constructed algorithm model is compared with the DBN proposed by Yang et al. the Convolutional Neural Network (CNN) proposed by Ullah et al. and the DBN-Hidden Markov Model (HMM) algorithm proposed by Xu et al. to analyse its performance. The recognition effects of each algorithm in the two datasets are analyzed. Results demonstrate that CNN developed by Ullah et al. has the worst accuracy on the KTH and UCF datasets, followed by DBN developed by Yang et al. and DBN-HMM developed by Xu et al. The constructed algorithm model can provide the highest accuracy, reaching about 90%, and the recognition accuracy of human sports behaviors of each algorithm on the KTH dataset is lower than that on the UCF dataset. On the KTH dataset, the recognition accuracy for boxing is the highest and running the lowest. Analyzing the model’s accuracy in the four scenes (S1, S2, S3, and S4) on the KTH dataset suggests that the recognition accuracy for the indoor scene (S4) is higher than that of the outdoor scenes (S1, S2, and S3). On the UCF dataset, the recognition accuracy for lifting is the highest, reaching 99%, and that for walking is the lowest, reaching 51%. Therefore, the proposed human sports recognition model can provide a higher accuracy than other classic DL algorithms, providing an experimental basis for subsequent human sports recognition research.