基于S_Kohonen神经网络的乒乓球运动识别

Design of ping-pong recognition based on S_Kohonen neural network

针对乒乓球运动识别方法通常无法实时识别、识别率低和识别算法复杂度高,从而导致穿戴式设备续航能力差等问题,提出一种基于遗传算法优化S_Kohonen(supervisedKohonen)神经网络的乒乓球运动实时识别方法,并完成系统设计.该系统通过单MPU6050六轴加速度传感器采集运动信号,采用动作端点检测算法提取动作始末端点,基于db4小波对动作信号进行3层分解,同时用遗传算法优化S_Kohonen神经网络对乒乓球常见的6种动作进行识别.实验结果表明:该运动识别方法离线平均识别率为99.17%,实时平均识别率为91.67%,待机功耗为0.28 mW,运行模式功耗为14 mW,识别时间为2 ms,证明该方法识别迅速、功耗低、识别准确率高.

In order to solve the problem of no real-time recognition,low recognition rate and poor endurance of wearable devices caused by high complexity of recognition algorithms in ping-pong recognition method,a real-time recognition method was proposed for ping-pong based on S_Kohonen(supervised Kohonen)neural network optimized by genetic algorithm,and the system design was completed.The single MPU6050 six-axis acceleration sensor was used to collect motion signal,and the start and end of the action was extracted by the endpoint detection algorithm.The motion signal was decomposed into three layers based on db4 wavelet,and the six common ping-pong motions was recognized by the S_Kohonen neural network optimized by algorithm optimization.The experiment result shows that the average offline recognition rate is 99.17%,the average real-time recognition rate is 91.67%,the standby power consumption is 0.28 mW,the power consumption of the operation mode is 14 mW and the recognition time is 2 ms.It proves that the method recognizes fast,and has low power consumption and high accuracy.