Human body posture estimation method based on bidirectional serialization modeling

Abstract

The invention discloses a human body posture estimation method based on bidirectional serialization modeling, which comprises the following steps: taking three continuous frames as input, fully utilizing the time sequence information of a video to calculate the approximate space range of each joint, and then returning to the specific position of each joint from a smaller range, therefore, the problems of inherent shielding, motion blur and the like in a human body posture estimation task are better solved, the generalization of the model is stronger, and the accuracy is higher. According to the invention, the time sequence information of the video is fully utilized, the reasoning capability of the model is enhanced, the key parts of the human body can be better estimated, and the invention has important significance in industries such as security and protection and short video platforms needing real-time posture extraction for analysis.

Description

technical field [0001] The invention belongs to the technical field of human body posture estimation, and in particular relates to a human body posture estimation method based on bidirectional serialization modeling. Background technique [0002] Human pose estimation is a cutting-edge research field in computer vision. Its goal is to locate key parts of the human body (such as wrists and ankles) in pictures or videos, so as to realize human pose estimation. Human body pose estimation is a bridge between communication machines and people. It has great practical significance and has been widely used in many fields, such as the field of stage animation. Real-time interactive animation effects can be generated by recognizing human gestures and actions; in the field of automatic driving, Car accidents can be avoided in advance by predicting the movement trend of pedestrians; in the security field, abnormal behavior can be detected by identifying specific gesture sequences. [0...

AI Technical Summary

Problems solved by technology

[0005] Traditional image-based pose estimation methods cannot effectively use these additional information, resulting in their inability to deal with situations such as highly entangled characters, mutual occlusion, and motion blur that often occur in video sequences, so it is difficult to achieve good results in video pose estimation.

In response to this problem, the literature [Flowing ConvNets for Human PoseEstimation in Videos-[CODE]–Pfister.T,Charles.J&Zisserman.A(ICCV 2015)] proposed to calculate the dense optical flow information between each two frames, and then use the flow-based time information to correct the initial pose estimation; when the optical flow can be calculated correctly, this method has achieved good results, but the calculation of optical flow is greatly affected by image quality, occlusion, etc., and it is impossible to accurately calculate all The optical flow information, and the calculation of optical flow information often requires a lot of computing power support

Some scholars also proposed to use Long Short-Term Memory (LSTM) to directly model the video to capture timing information. However, due to the structural limitations of the LSTM network itself, this method can only be obtained when the characters in the video frame are relatively sparse. Good effect, when used in complex scenes, it still cannot deal with occlusion, motion blur, etc.

Images