Lightweight human body posture estimation method and system based on streaming attention

Abstract

The invention discloses a lightweight human body posture estimation method and system based on streaming attention. The method comprises the steps that a target image needing posture estimation is acquired; carrying out standardized clipping and normalization processing on the target image; inputting the target image into a trained lightweight human body posture estimation network based on streaming attention, and outputting a posture estimation heat map result; a model training step: performing down-sampling on training images in the training data set through convolution of different scales, obtaining four first feature maps of different scales as input of an up-sampling module through multiple groups of ghost lightweight modules with streaming attention, and performing convolution respectively to obtain four second feature maps of the same number of channels; and carrying out bilinear up-sampling operation to obtain a double feature map, carrying out feature fusion, outputting a third feature map, and converting to obtain a heat map required by human body posture estimation. The method is high in recognition precision and low in complexity.

Description

technical field [0001] The present invention relates to the field of human body posture estimation, in particular to a light-weight human body posture estimation method and system based on streaming attention. Background technique [0002] In the direction of computer vision, human pose estimation is a very hot and important topic. Its main task is to obtain human joint points and limb information based on images or videos. After obtaining this information, some cognitive tasks can be further completed. Such as behavior recognition and human-computer interaction. [0003] There are generally two ideas for human pose estimation tasks: top-down and bottom-up, so the structure design of the neural network is also of these two ideas. The top-down human pose estimation algorithm consists of two steps: object detection and single-person pose estimation. That is, first use the classic target detection algorithm such as Faster-RCNN to frame the people in the image, and then perfo...

AI Technical Summary

Problems solved by technology

[0005] The present invention provides a lightweight human body pose estimation method and system based on streaming attention, which is used to solve the technical problems of low recognition accuracy and high complexity of the existing human body pose estimation network

Images