Research Of Efficient Neural Network Algorithms For Human Skeleton Analysis

Deep learning has become an upsurge of big data and artificial intelligence since it came out.Models,which are composed of deeper structure and complex connection,show powerful performance on various deep learning topics.However,the growth of hardware computing resources can not meet the explosive growth of model size and computation.Therefore,how to design network structure and improve computational efficiency is an important research direction of deep learning.For domain-specific problems,domain-related knowledge can be integrated into algorithms or network structures,which provides inspiration for designing efficient algorithms.Based on it,this thesis focuses on three important components of human skeleton analysis:object detection,image classification and keypoint localization.Aiming at the problem of exhaustive ROI cost in object detection,performance is improved by integrating prior knowledge into the algorithm and using extended data representation range.Aiming at the problem that image classification relies too much on the black box feature extraction ability of depth neural network,the tree neural network is constructed by integrating human hierarchical semantic knowledge.Aiming at the problem of overly relying on stacked hourglass models in keypoint localization,on the basis of reducing the convolution complexity,the relationship of human skeleton is used to train 2d and 3d keypoint datasets together to enhance the generalization ability of the model and improve the accuracy.For each task,experiments are carried out on the relevant mainstream datasets.The proposed algorithm can maintain or outperform mainstream methods in accuracy,and significantly reduce the time-consuming.The experiment show that the deep convolutional neural network can combine the relevant domain knowledge to train and inference effectively,and application scenarios have been further expanded.