基于改进的YOLOv3实现手势识别的人机交互方法

Human-Computer Interaction Method for Gesture Recognition Based on Improved YOLOv3

随着人工智能技术的飞速发展,人机交互方法也发生了巨大的变化.鉴于目前主要的人机交互方式仍是使用键盘、鼠标和触控板组合的传统交互方式,本文提出一种基于改进的YOLOv3实现手势识别的人机交互方法,通过Kmeans对标签的边界框进行聚类,然后运用Mosaic数据增强丰富小目标,最后采用自定义最小化边界框中心点距离的GCDIoU损失函数优化模型参数.在自建数据集上进行实验验证,该模型针对手势小目标的检测准确率达到98.87%,召回率达到99.98%.结果表明,Mosaic数据增强应用于小目标检测具有很好的效果,而GCDIoU损失函数则加快了模型的收敛.

With the rapid development of artificial intelligence technology,human-computer interaction has also undergone tremendous changes.In view of the fact that most people are still using the traditional interaction method of keyboard,mouse and touchpad,in this article we propose a method of real-time gesture recognition and human-computer interaction based on the improved YOLOv3.Specifically,the bounding box of the label was first clustered through K-means.Then the small targets were enriched by Mosaic data enhancement,and finally the GCDIoU loss function that minimizes the distance between the center points of the bounding box was used.Through the above methods,the final model's detection accuracy for small gesture targets reached 98.87%,and the recall rate reached 99.98%.Therefore,Mosaic data enhancement has a good effect when applied to small target detection,while GCDIoU loss function speeds up the convergence of the model,and the effect is better.