为了方便观众更好地在观看比赛直播和录像时理解裁判手势的含义, 或帮助录像分析师分析比赛视频, 设计了一种实时篮球裁判手势检测与识别系统Yolov5-BR(Yolov5-Basketball Referee). 首先, 采用目标检测中的Yolov5 算法为基础模型, 对其边界框的交并比(intersection over union, IoU) 损失函数完全交并比(complete intersection over union, CIoU) 进行加权处理, 增强预测框的鲁棒性; 其次, 在C3 模块后加入注意力机制, 产生更具分辨性的特征表示, 从而提升网络识别性能; 此外, 在检测层头部融入自适应特征融合机制, 充分利用图像高层语义信息; 最后, 对目标置信度损失函数进行不对等加权处理, 从而提高对小目标检测的鲁棒性. 在自制的裁判手势数据集上, Yolov5-BR 取得了95.4% 的mAP 值, 本地视频检测速率为55.5 帧/s, 外接摄像头分辨率为$1 280\times 960$, 检测速率为25 帧/s. 实验结果表明, Yolov5-BR 相对于原始模型在检测裁判手势的性能上有所提升, 保持了较高的准确率、稳定性与实时性.

To help the audience better understand the meaning of the referee's gesture when watching a live broadcast or a video of a basketball game, or to help video analysts analyze the game video, a real-time basketball referee gesture detection and recognition system, Yolov5-Basketball Referee (Yolov5-BR), was designed. The Yolov5 target detection algorithm was used as the basic model, and the intersection over union (IoU) and complete IoU (CIoU) loss functions of its boundary box were weighted to enhance the robustness of the prediction box. After the C3 module, an attention mechanism was added to generate more distinguishing feature representation and improve the network recognition performance. In addition, an adaptive feature fusion mechanism was incorporated into the head of the detection layer to make full use of the high-level semantic information of the image. Finally, the target confidence loss function was weighted unequally to improve the robustness of small-target detection. On a self-made referee gesture dataset, Yolov5-BR achieved a 95.4% mAP value, with a local video detection rate of 55.5 frame/s, an external camera resolution of $1 280 \times 960$, and a detection rate of 25 frame/s. Experimental results show that, compared with the original model, Yolov5-BR can effectively improve the performance of judging gestures while maintaining high accuracy, stability, and real-time response.