3D posture estimation method based on multi-view deep sensor frame

Abstract

The invention provides a 3D posture estimation method based on a multi-view deep sensor frame. The method mainly comprises a specific view sensor network, a multiple view integration network, layeredjump connection, data preprocessing, training and evaluation,; and the specific view sensor network extracts 2D shapes and layered texture information from different views, and generates mapping for the different views, a hourglass network composed of an encoder and a decoder form a high-resolution heat image of each joint, jump connection of the hourglass network is used to realize the multiple view integration network, the multiple view integration network integrates information from all available views, and an accurate 3D posture is provided. Hierarchical information is combined with estimated joint heat images to deduce the 3D dstructure, limitation of a direction measurement and observation system can be overcome, and the 3D posture is estimated more accurately.

Description

technical field [0001] The invention relates to the field of posture estimation, in particular to a three-dimensional posture estimation method based on a multi-view deep sensor framework. Background technique [0002] Human pose estimation is a research hotspot in the field of computer vision in recent years. The computer system extracts the human body posture picture from the image or video, analyzes and counts the human body posture, and then judges the character's behavior. Therefore, human pose estimation has extremely wide applications. In the application of the abnormal behavior detection system, through the real-time detection and analysis of the person's posture in the video surveillance screen, when there are fighting, stealing and other behaviors in the screen, the system can record and issue an alarm in time. In the application of sports posture analysis, the video analysis system of three-dimensional posture estimation is used to analyze and process the video ...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problems that the existing systems are relatively expensive, time-consuming, and inaccurate, the purpose of the present invention is to provide a 3D pose estimation method based on a multi-view deep perceptron framework. The two-dimensional shape and layered texture information are extracted from the image, the specific view perceptron network generates a map for each view, the hourglass network composed of encoder and decoder constructs a high-resolution heat map of each joint, and the skip connection of the hourglass network is used to Implement a multi-view integration network that synthesizes information from all available views to provide accurate 3D pose

Images