Human body key point detection method based on multi-scale cascaded hourglass network

Abstract

Human body key point detection method based on multi-scale cascaded HourGlass network: 1: Generate 128*128, 64*64 and 32*32 feature maps from the input image; 2: Send 128*128 feature maps to the first 2 Modified HourGlass In the module, the input and output of these two modules are both 128*128 feature maps; Three: Downsample the 128*128 feature map and 128*128 heat map output by the second Modified HourGlass module into a 64*64 feature map and The 64*64 feature map generated in step 1 is sent to the middle 4 Modified HourGlass modules. The input and output of these 4 modules are both 64*64 feature maps; 4: The 64*64* output from the sixth Modified HourGlass module The 64 feature map and the 64*64 heat map are down-sampled into a 32*32 feature map and the 32*32 feature map generated in step 1 is sent to the last 2 modules; 5: Extract the loss function of 8 Modified HourGlass modules, Add the loss function; the network performs backward feedback according to the loss function, and stops training when the loss function reaches the expected value and stabilizes.

Description

technical field [0001] The invention belongs to the technical field of single-person pose estimation, and in particular relates to a human body key point detection method based on a multi-scale cascaded HourGlass network. Background technique [0002] Classic single-person pose estimation methods are based on graph-structured models. In this model, the spatial correlation of body parts is represented as a tree-structured graph model. Later, in order to enhance the tree structure to capture the symmetry and long-distance relationship between the key points of the human body, a non-tree model was developed by introducing a loop mechanism. Overall, single-person pose estimation based on graph-structured models has shortcomings such as repeated counting of key points and complex modeling. [0003] With the development of deep learning, many CNN-based single-person pose estimation methods have emerged. For example, DeepPose directly returns the coordinates of the joints. The ...

AI Technical Summary

Problems solved by technology

Although the HourGlass network has a high detection rate for human key points on data sets such as MPII, its structure is still unreasonable and does not make full use of the rich information contained in the feature map in the network.

Images