《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (7): 2173-2181.DOI: 10.11772/j.issn.1001-9081.2022060810

• 人工智能 • 上一篇    下一篇

轻量化篮球裁判手势识别算法

李忠雨1, 孙浩东1, 李娇1,2,3()   

  1. 1.上海大学 微电子研究与开发中心, 上海 200444
    2.上海大学 机电工程与自动化学院, 上海 200444
    3.上海大学 新型显示技术及应用集成教育部重点实验室(上海大学), 上海 200444
  • 收稿日期:2022-06-06 修回日期:2022-09-07 接受日期:2022-09-09 发布日期:2023-07-20 出版日期:2023-07-10
  • 通讯作者: 李娇
  • 作者简介:李忠雨(1997—),男,重庆人,硕士研究生,主要研究方向:目标检测;
    孙浩东(1998—),男,山西大同人,硕士研究生,主要研究方向:目标检测;
    李娇(1975—),女,上海人,讲师,博士,主要研究方向:模式识别。
  • 基金资助:
    国家自然科学基金资助项目(52107239)

Lightweight gesture recognition algorithm for basketball referee

Zhongyu LI1, Haodong SUN1, Jiao LI1,2,3()   

  1. 1.Microelectronic Research and Development Center,Shanghai University,Shanghai 200444,China
    2.School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200444,China
    3.Key Laboratory of Advanced Display and System Applications,Ministry of Education (Shanghai University),Shanghai 200444,China
  • Received:2022-06-06 Revised:2022-09-07 Accepted:2022-09-09 Online:2023-07-20 Published:2023-07-10
  • Contact: Jiao LI
  • About author:LI Zhongyu, born in 1997, M. S. candidate. His research interests include object detection.
    SUN Haodong, born in 1998, M. S. candidate. His research interests include object detection.
    LI Jiao, born in 1975, Ph. D., lecturer. Her research interests include pattern recognition.
  • Supported by:
    National Natural Science Foundation of China(52107239)

摘要:

针对一般手势识别算法的参数量、计算量和精度难以平衡的问题,提出一种轻量化篮球裁判手势识别算法。该算法在YOLOV5s (You Only Look Once Version 5s)算法的基础上进行重构:首先,用Involution算子替代CSP1_1的卷积算子,以扩大上下文信息捕获范围并减少核冗余;其次,在C3模块后加入协同注意力(CA)机制,以得到更强的手势特征提取能力;然后,用轻量化内容感知上采样算子改进原始上采样模块,并将采样点集中在目标区域而忽略背景部分;最后,利用以SiLU作为激活函数的Ghost-Net进行轻量化剪枝。在自制的篮球裁判手势数据集上的实验结果表明,该轻量化篮球裁判手势识别算法的计算量、参数量和模型大小分别为3.3 GFLOPs、4.0×106和8.5 MB,与YOLOV5s算法相比,分别减少了79%、44%和40%,mAP@0.5为91.7%,在分辨率为1 920×1 280的比赛视频上的检测帧率达到89.3 frame/s,证明该算法能满足低误差、高帧率和轻量化的要求。

关键词: 目标检测, 手势识别, Involution算子, 注意力机制, 上采样, Ghost-Net

Abstract:

Aiming at the problem that the number of parameters, calculation amount and accuracy of general gesture recognition algorithms are difficult to balance, a lightweight gesture recognition algorithm for basketball referee was proposed. The proposed algorithm was reconstructed on the basis of YOLOV5s (You Only Look Once Version 5s) algorithm: Firstly, the Involution operator was used to replace CSP1_1 (Cross Stage Partial 1_1) convolution operator to expand the context information capturing range and reduce the kernel redundancy. Secondly, the Coordinate Attention (CA) mechanism was added after the C3 module to obtain stronger gesture feature extraction ability. Thirdly, a lightweight content aware upsampling operator was used to improve the original upsampling module, and the sampling points were concentrated in the object area and the background part was ignored. Finally, the Ghost-Net with SiLU (Sigmoid Weighted Liner Unit) as the activation function was used for lightweight pruning. Experimental results on the self-made basketball referee gesture dataset show that the calculation amount, number of parameters and model size of this lightweight gesture recognition algorithm for basketball referee are 3.3 GFLOPs, 4.0×106 and 8.5 MB respectively, which are only 79%, 44% and 40% of those of YOLOV5s algorithm, mAP@0.5 of the proposed algorithm is 91.7%, and the detection frame rate of the proposed algorithm on the game video with a resolution of 1 920×1 280 reaches 89.3 frame/s, verifying that the proposed algorithm can meet the requirements of low error, high detection rate and lightweight.

Key words: object detection, gesture recognition, Involution operator, attention mechanism, upsampling, Ghost-Net

中图分类号: