Human body posture estimation method based on key point correlation

Abstract

The invention belongs to the field of human body posture estimation, and particularly relates to a human body posture estimation method based on key point correlation. The method comprises the following steps: acquiring image data in real time, and preprocessing the acquired image data; inputting the preprocessed image data into a trained human body posture estimation model based on key point correlation to obtain a human body posture estimation result; according to the method, key point groups based on strong correlation are constructed, relation information of key points with strong correlation is fused by using a relation network, and correlation loss and correlation confusion penalty loss are constructed, so that feature fusion between the key points is effectively enhanced, the confusion problem between the key points is improved, the prediction precision is further improved, and the prediction efficiency is improved. Estimation is more accurate, and the method has good economic benefits.

Description

technical field [0001] The invention belongs to the field of human body posture estimation, and in particular relates to a human body posture estimation method based on the correlation of key points. Background technique [0002] Human pose estimation refers to the process of detecting the positions of human joints in pictures or videos and labeling the categories of joints. This process is also called human keypoint detection. The field initially mostly adopts graph structure-based methods to deal with the human pose problem. Therefore, the deformable component model based on the graph structure method came into being. Since these methods can only build a model for all the subsets that are in contact between the joints of the human body, although high efficiency is obtained, it is affected by the occlusion factor The impact is large and the ability is limited. [0003] In recent years, thanks to the rapid development of the field of deep learning, computer learning of ima...

AI Technical Summary

Problems solved by technology

However, these methods only focus on network feature extraction, but ignore the feature information relationship of the human body joints themselves, resulting in confusion between two key points due to unclear semantics in dark and occluded scenes (such as people in Difficulty distinguishing left and right shoulders when turned sideways)

[0006] In the prior art, a group training method based on the correlation of key points of the human body divides key points with strong correlation into a group for common training. Although this method can solve the above problems, it also ignores the global semantics of the human body as a whole. Problems such as information fusion between information and key points, and two irrelevant or weakly correlated key points in the same group may cause the network to show a negative migration trend, making the predicted key points more likely to be confused

Images