Human Body Attitude Estimation Based On Convolution Neural Network

In order to improve the accuracy of human attitude estimation by convolution neural network in two-dimensional images,the problem of searching joint space in the process of human attitude estimation was discussed.In this paper,a prior distribution based joint search space reduction method was proposed,and previous researchers use convolutional neural networks to build joint appearance models that only consider the local features of images.In this paper,an improved convolution neural network model was proposed,which ignores the problem that human beings depend on both local and global features while recognizing objects,and that the features extracted from different convolution kernels are treated equally.And make human gestures on FLIC and LSP datasets.The experimental results show that the proposed method can effectively reduce the joint space and improve the accuracy of attitude estimation.Considering that human posture is dynamic in most cases,convolutional neural networks are applied to the recognition and estimation of dynamic human posture in video.Firstly,the information in 3D video should be converted into two-dimensional feature image.Based on the common motion foreground detection algorithm,which is difficult to obtain complete contour image,a human body silhouette extraction algorithm based on DOG image is proposed.The integrity of feature contour was improved effectively in the process of attitude estimation.Finally,the improved convolution neural network proposed in Chapter 3 is used to train and recognize the two-dimensional feature map,which effectively improves the attitude estimation.Accuracy rate.Based on the three-dimensional convolutional neural network model,through the experiment based on contour feature and motion feature,the best two-dimensional feature graph combination for experiment was obtained,namely "optical flow graph-frame difference graph-3 frame difference graph".Using the network model with convolution kernel size of 7 × 7 and convolution layer size of 10,the accuracy of convolution kernel is up to 92%.Then the database was processed with data equalization,and the accuracy of human attitude estimation was 94.8%.Compared with the unbalanced data,the accuracy was increased by 2.8%.Experimental results show that 3D convolution god Compared with the traditional methods,the network not only reduces the workload,but also has good robustness.Figure 40;Table 20;Reference 48.