基于小样本手部关键点的MLP网络提升3D光场交互准确度方法

Method for improving the accuracy of 3D light field interaction based on a small dataset of hand key points using an MLP network

针对当前3D光场手势交互存在识别率低、识别速度慢、深度学习网络需要较多数据样本的问题,本文提出了一种基于小样本手部关键点的多层感知器(Multi-Layer Perceptron,MLP)网络提升3D光场交互准确度方法,识别速度达到毫秒级。在手部关键点采集过程中,从不同位置采集得到的同一种手势关键点三维数据存在显著差异。为了消除差异,本文提出在同一右手笛卡尔坐标系下,通过位移和罗德里格旋转公式对简化后的手势模型进行位姿变换,将同一种手势归一化。一个MLP神经网络被用来从归一化后的手部关键点跳变关系中提取手部特征。实验结果表明,本文提出的方法对3D光场交互中的简单手势识别率为95%以上,对复杂手势的识别率为90%以上。与此同时,该方法在小样本数据集训练下表现出优秀的性能,能够满足精确和快速手势识别的要求。最后,本文展示了一种将所提出的方法成功应用于3D光场交互的场景。

To address the issues of low recognition rate,slow recognition speed,and the need for large amounts of data samples in current 3D light field gesture interaction,this paper proposes a method based on a small dataset of hand key points using a multi-layer perceptron(MLP)network to improve the accuracy of 3D light field interaction,with recognition speed reaching the millisecond level.In the process of collecting hand key points,there are significant differences in the three-dimensional data of the same type of hand gesture collected from different locations.In order to eliminate these differences,this paper proposes a method of normalizing the same gesture through pose transformation of the simplified gesture model in the same right-hand Cartesian coordinate system using displacement and Rodrigues rotation formula.An MLP neural network is utilized to extract hand features from the normalized hand key points transition relationships.Experimental results show that the proposed method has a recognition rate of above 95%for simple gestures in 3D light field interaction,and a recognition rate of above 90%for complex gestures.Furthermore,the proposed method demonstrates excellent performance under training with a small dataset,meeting the requirements of both accurate and fast gesture recognition.Finally,this paper presents a successful application of the proposed method to a 3D light field interaction scenario.