Precise three-dimension hand and human body attitude estimating method based on single depth image

Abstract

The invention provides a precise three-dimension hand and human body attitude estimating method based on a single depth image; the method mainly comprises the following contents: a network model, an improved target position, system input, a voxel-to-voxel prediction network; the method comprises the following steps: firstly providing a network whole framework, using a convolution nerve network-based method to improve the target position, using a back projection technology to build the system input, finally using four structure blocks and an encoder and a decoder to form the voxel-to-voxel prediction network, wherein the four blocks comprises a volume basic block, a volume residual block, a volume lower sampling block and a volume upper sampling block. The method can solve the perspective distortion and nonlinear mapping problems, can obtain high precision three-dimension hand and human body attitude estimations, uses less time consumption, and can realize real time human body behaviorprediction and estimation.

Description

technical field [0001] The invention relates to the field of three-dimensional hand and human posture estimation, in particular to an accurate three-dimensional hand and human posture estimation method based on a single depth image. Background technique [0002] Human behavior interaction means that the computer understands and responds to human actions and behaviors by locating and identifying humans, tracking human limb movement trajectories, and tracking facial expression features. Its application background is very extensive, mainly focusing on human-computer interaction, virtual reality, smart home, smart security, smart video surveillance, patient monitoring system, auxiliary training for athletes, and content-based video retrieval and smart image compression. The method of human behavior interaction. For example, by detecting and estimating suspicious hand movements or postures of people in public places such as train stations and airports, it can assist security per...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problems of perspective distortion and nonlinear mapping, the purpose of the present invention is to provide an accurate 3D hand and human pose estimation method based on a single depth image. Firstly, the overall architecture of the network is given, and then the convolutional neural network The position of the target is improved, and then the input of the system is constructed using the back-projection technique, and finally the voxel pair is composed of four types of building blocks: the volume basic block, the volume residual block, the volume down-sampling block and the volume up-sampling block, as well as the encoder and the decoder. Voxel prediction network

Images