基于Mediapipe的手部姿态数据集制作方法

Dataset creation method for hand pose estimation based on Mediapipe

传统的手势姿态检测方法存在着数据量大、无效特征多、标注数据需求高等缺陷,由于个体差异和定制化需求,需要识别的手部动作不尽相同。此外,某些特殊场景下,基于公开数据集制作的模型并不能准确判断手部姿态。谷歌发布的Mediapipe手部检测模型通过已训练成熟的手部关键点检测算法,可以直接获取手部关键点信息。因此,提出了一种基于Mediapipe模型的便捷的训练集收集程序,能够使用该数据集训练固定场景下的实时手势姿态检测算法。这种方法不仅提高了准确率,还减少了数据量和训练时间,从而提高了算法的效率和可靠性。同时建立了10分类手势数字数据集,通过多层感知机进行机器学习,在识别率和灵敏度上取得了较好的成果,正确率达93.38%.

Traditional hand gesture recognition methods have drawbacks such as large data requirements,many ineffective features,and high demand for annotated data.Additionally,the required hand movements to be recognized vary according to individual differences and customized needs.Furthermore,models trained on public datasets may not accurately predict hand pose in some special scenarios.At present,Google’[KG-*3]s Mediapipe hand detection model can directly obtain hand keypoint information through a well-trained algorithm.In this paper,we propose a convenient training dataset collection program based on the Mediapipe model,which can be used to train real-time hand pose detection algorithms in fixed scenes with low computational and resource costs.This approach not only improves accuracy,but also reduces data requirements and training time,so as to enhances the efficiency and reliability of the algorithm.Meanwhile,we also established a 10,classified gesture digital data set and carried out machine learning by multi-layer perceptron,achieving good results in recognition rate and sensitivity,with an accuracy rate of 93.38%.