基于MTCNN的坐姿行为识别

Sitting behavior recognition based on MTCNN

针对长时间处于不正确坐姿下引发的颈椎疾病、近视发病率增高问题,提出一种基于MTCNN的坐姿行为识别方法。将MTCNN算法应用到坐姿识别中检测人脸关键点,根据肩膀ROI区域,利用肩膀轮廓多边形近似算法定位肩膀位置;通过卷积神经网络和肩膀定位获得坐姿关键点,计算坐姿特征构成高维特征向量。SVM分类实验结果表明,该方法的坐姿行为识别正确率达到95.9%,与其它算法相比,其能够准确识别坐姿行为,通过增添肩膀特征,识别率提高了1.8%。

To solve the problem of increased disease rate of cervical spondylosis and myopia caused by prolonged sitting in an incorrect posture,the way of sitting behavior recognition based on joint face detection and alignment using multi-task cascaded convolutional networks(MTCNN)was proposed.MTCNN algorithm was used to detect the key points of human face in sitting position recognition,and the shoulder position was located using the shoulder contour polygon approximation algorithm according to the shoulder region of interest.The key points of sitting posture were collected using convolution neural network and shoulder position.The sitting posture features were calculated and the high dimensional eigenvectors were generated.Experimental results of support vector machine(SVM)classification show that the correct rate of the method proposed reaches 95.9%,which can accurately recognize sitting posture behaviors compared with other algorithms.The recognition rate is increased by 1.8%with shoulder features added.