低成本惯性传感器的信号增强与手势识别

Signal Enhancement and Gesture Recognition for Low⁃Cost Inertial Sensors

针对低成本惯性传感器中普遍存在的严重随机误差,提出了一种三阶段信号增强策略,通过复杂网络确定性指标,首次实现了对IMU信号随机性的量化,进而利用经验模态分解、小波变换对随机误差进行有效控制。进一步,首次在IMU运动识别任务中将时频特征与复杂网络拓扑特征相结合,在不同用户运动习惯各异和IMU型号多样且成本低廉的情况下,将62种三维手势动作的识别精度提升至92.4%。此外,对比了多种机器学习与深度学习模型在手势识别任务上的精度,进一步论证了复杂网络拓扑特征在运动状态识别、时间序列分析领域的重要价值。

The low⁃cost inertial sensors have serious errors.To address the problem,a three⁃stage signal enhancement strategy is pro⁃posed.For the first time,the IMU signal randomness is quantified through using the complex network deterministic index.Then,random errors can be controlled effectively by empirical modal decomposition and wavelet transform.Furthermore,the time⁃frequency features are fused with the topology features of the complex network in the IMU motion recognition task.For the first time,the motion habits of different users and various types of IMU with different performances are considered,in which case,the recognition accuracy for 62 kinds of 3D gestures is improved to 92.4%.In addition,the value of complex network topological features in motion state recognition and time series analysis is demonstrated by comparing the gesture recognition accuracy of various machine learning and deep learning models.