Aiming at the interference of the visual pose estimation method, such as cover and occlusion, a method of estimating human three-dimensional sitting posture based on the seat surface pressure image is proposed. The cross-domain relationship between seat surface pressure distribution and human three-dimensional posture is established. A posture training system based on pressure and vision is designed. The array pressure sensor is embedded in the seat surface to perceive the posture, and the time stamp is used to realize the synchronization of the visual image matching with the binocular camera. Bilateral filtering is used to eliminate the peak noise of pressure images. Nineteen 3D keypoints are extracted from binocular vision images by OpenPose estimation and triangulation. To improve the accuracy of attitude estimation, a stochastic gradient descent method to minimize the loss function is proposed to optimize the coordinates of 3D keypoints. The 3D confidence graph of keypoints is further generated by 3D Gaussian filter. A multi-layer convolutional neural network pressure-vision cross-domain deep learning model is formulated. Continuous multi-frame pressure images are used as input of the model, and 3D pose estimation results of 3D key point coordinates and their confidence graphs are used as supervision. Based on the pressure distribution of the array sensor on the chair surface, the algorithm can accurately estimate the 3D sitting posture including 19 human key points. The average error of 19 key points is 9. 7 cm on the verification set.