Human body posture estimation method based on joint relation

Abstract

The invention relates to a human body posture estimation method based on a joint relation, and the method comprises the following steps: S1, constructing a joint relation module which comprises two sub-modules: a channel-based feature relation module and an adjacent joint space relation module; s2, constructing a human body posture estimation model based on a joint relation; s3, training a human body posture estimation model based on a joint relation by utilizing the marked human body posture data; and S4, performing a human body posture estimation task based on a single image by using the trained human body posture estimation model added with the joint relation module to obtain a predicted human body posture. Compared with the prior art, the invention effectively overcomes the problem that the positions of four-limb joints with high degree of freedom, such as wrist joints, ankle joints and shielded invisible joints, are difficult to detect in the image, and is high in human body posture estimation accuracy.

Description

technical field [0001] The invention relates to the field of human body pose estimation, in particular to a method for estimating human body pose based on joint relationships. Background technique [0002] Human pose estimation is a traditional task in the field of computer vision. Human pose estimation includes the detection of human key points and the generation of human poses. The "key points" in human key point detection refer to important joints such as the top of the human body, shoulders, elbow joints, wrist joints, and ankle joints. The generated human body posture is the complete human skeleton information. With the innovation of computer vision technology, human pose estimation has also gone through a process from manually extracting features to using deep convolutional neural networks as a tool. In recent years, the development of basic deep convolutional neural network structure and performance has also greatly improved the level of extracting human joint featu...

AI Technical Summary

Problems solved by technology

This approach cannot be trained end-to-end, and the graph convolutional network stage is computationally intensive

[0005] The human body is a non-rigid body, and the rotation of each joint has a great degree of freedom. The free rotation of multiple joints can be superimposed on each other. The joints at the far end of the limbs, such as wrist joints and ankle joints, also have multiple degrees of freedom. The location changes, which increases the difficulty of prediction, but the existing methods do not pay attention to the difference in the detection difficulty of different types of joint points brought about by the special structure of the human body; at the same time, in the multi-person pictures that are closer to the real scene, due to Interaction with people, the diversity of positional relationships and the complexity of life scenes, the position of key points of the human body and body information are all facing serious occlusion problems

The occluded human joints are also part of the human pose, and the existing methods do not take targeted optimization measures for the invisible joints that are more difficult to detect correctly

Images