Human 3D Pose Estimation Based on Sequence Graph Convolution: IEEE ITAIC(ISSN:2693-2865)

Thesis mainly studies the two-stage 3D pose estimation method of human body with 2D keypoint sequence as input. The sequence graph structure is established by intra-frame connection of different joint points in the input sequence and inter-frame connection of the same joint point. Then the graph convolutional network and the dilated temporal convolutional network structure are used to obtain the spatial and temporal features of the sequence spatiotemporal graph to estimate the 3D coordinate estimation of the human pose. Make full use of the feature of dilated temporal convolutional network to gradually reduce the number of input sequence frames to reduce the volume of the high-level sequence spatiotemporal graph structure to reduce the training time and improve the network performance.