CN-121982200-A - Real-time human body reconstruction method for nerve symbol distance field and electronic equipment

CN121982200ACN 121982200 ACN121982200 ACN 121982200ACN-121982200-A

Abstract

The invention provides a real-time human body reconstruction method of a nerve symbol distance field and electronic equipment, and relates to the technical field of artificial intelligence. The method comprises the steps of constructing and training a gesture conditional neural network model, inputting a query point into the trained target gesture conditional neural network model, outputting a symbol distance field predicted value corresponding to the query point, wherein a double encoder module in the gesture conditional neural network model extracts surface features and gesture features according to human body information of a target individual, a feature fusion module carries out feature fusion on the surface features and the gesture features to obtain a condition vector, and a FiLM condition decoding module outputs the symbol distance field predicted value corresponding to the query point according to the query point and the condition vector. The invention adopts a double-branch encoder architecture to extract and fuse the multi-modal characteristics of sparse input data, realizes the depth fusion of geometric information and attitude information, and combines FiLM linear modulation, thereby having high reconstruction precision, high speed and low input dependence.

Inventors

LIU JIE
MA JUNQIANG
LI YUE
LIU ZIAN
Jin Chencong

Assignees

河北科技工程职业技术大学

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. A method for real-time human reconstruction of a neuro-symbol distance field, comprising: Acquiring human body information of a target individual; constructing a gesture conditional neural network model according to the human body information of the target individual; Training the attitude condition neural network model to obtain a target attitude condition neural network model after training; inputting a query point into the trained target attitude condition neural network model, and outputting a symbol distance field predicted value corresponding to the query point; The gesture conditional neural network model comprises a double encoder module, a feature fusion module and a FiLM conditional decoding module, wherein the double encoder module is used for extracting surface features and gesture features according to human body information of a target individual, the feature fusion module is used for carrying out feature fusion on the surface features and the gesture features to obtain a conditional vector, and the FiLM conditional decoding module is used for outputting a symbol distance field predicted value corresponding to a query point according to the query point and the conditional vector.
2. The method for reconstructing human body in real time according to claim 1, wherein the human body information of the target individual comprises a human body skeleton key point coordinate vector of the target individual and a human body surface sparse sampling point coordinate vector of the target individual; the key point encoder is used for encoding the coordinate vector of the key point of the human skeleton of the target individual to obtain the gesture feature; the surface point encoder is used for encoding the coordinate vector of the sparse sampling point on the surface of the human body of the target individual to obtain the surface characteristics.
3. The method for reconstructing human body in real time by using a neural symbol distance field according to claim 2, wherein the surface point encoder comprises three 1D convolution layers and a maximum pooling layer which are sequentially connected; The key point encoder comprises three full-connection layers which are connected in sequence.
4. The method for reconstructing a human body in real time for a nerve symbol distance field according to any one of claims 1 to 3, wherein the feature fusion module comprises a first splicing layer, a fusion MLP layer and a second splicing layer which are sequentially connected; the first splicing layer is used for splicing the surface features and the gesture features to obtain joint features; The fusion MLP layer is used for carrying out feature fusion and nonlinear mapping on the combined features to obtain hidden codes; And the second splicing layer is used for splicing the hidden code and the gesture feature to obtain the condition vector.
5. The method for real-time human reconstruction of a neurosymbol distance field according to any one of claims 1-3, wherein said FiLM condition decoding module comprises a FiLM parameter generator and a FiLM condition decoder; The FiLM parameter generator is used for generating a scaling parameter and an offset parameter according to the condition vector; The FiLM condition decoder is used for decoding the query point according to the scaling parameter and the offset parameter to obtain a symbol distance field predicted value corresponding to the query point.
6. The method for real-time human reconstruction of a neurosymbol distance field according to claim 5, wherein said FiLM parameters generator comprises two MLP layers connected in sequence; the FiLM conditional decoder comprises an input projection layer, a plurality of FiLM modulation layers and an output head which are connected in sequence.
7. A method for real-time human body reconstruction of a neural symbol distance field according to any one of claims 1 to 3, wherein training the gesture conditional neural network model to obtain a trained target gesture conditional neural network model comprises: constructing a multi-target joint loss function, and training the attitude condition neural network model by adopting a near-surface area training sample, a middle-distance area training sample and a uniform sampling area training sample to obtain a target attitude condition neural network model after training; the distance between the near-surface region training samples is smaller than a first preset distance, and the distance between the intermediate-distance region training samples is not smaller than the first preset distance and smaller than a second preset distance.
8. The method of claim 7, wherein the multi-objective joint loss function The method comprises the following steps: Wherein, the Is Huber loss; Classifying cross entropy loss for symbols; Is a loss of the distance between two-way chamfer angles; Regularizing the term for the hidden code; 、、 And Is a coefficient.
9. The method of real-time human reconstruction of a neuro-symbol distance field as claimed in claim 8, 。
10. An electronic device comprising a memory and a processor, the memory storing a computer program, the processor implementing the neural symbol distance field real-time human reconstruction method of any one of claims 1 to 9 when the computer program is executed.

Description

Real-time human body reconstruction method for nerve symbol distance field and electronic equipment Technical Field The invention relates to the technical field of artificial intelligence, in particular to a real-time human body reconstruction method for a nerve symbol distance field and electronic equipment. Background In Human-computer collaboration (Human-Robot Collaboration, HRC), real-time accurate Human body geometric modeling is a key technical basis for achieving secure interaction. The symbol distance field (SIGNED DISTANCE FIELD, SDF) is used as a continuous implicit geometric representation method, can return the nearest distance from any space query point to the surface of the human body and the inner and outer symbols thereof, and provides an efficient mathematical framework for collision detection and motion planning of the robot. In a dynamic human-computer collaboration scenario, real-time human SDF reconstruction is a core challenge facing SDF modeling. In the prior art, the real-time human SDF reconstruction method comprises a voxelized SDF method, a geometric approximation method, a neural implicit method and the like. The voxelized SDF method is poor in instantaneity and insufficient in precision, the geometric approximation method is high in calculation speed, geometric details of complex curved surfaces of a human body cannot be accurately captured, dense surface points are needed to be used as input in the neural implicit method, and the method is difficult to be directly applied to a real-time man-machine cooperation scene. The existing real-time human SDF reconstruction method cannot meet the requirements of high precision, real-time performance and sparse input at the same time, and cannot meet the actual application requirements. Disclosure of Invention The embodiment of the invention provides a real-time human body reconstruction method of a nerve symbol distance field and electronic equipment, which are used for solving the problem that the existing real-time human body SDF reconstruction method cannot meet the requirements of high precision, real-time performance and sparse input at the same time. In a first aspect, an embodiment of the present invention provides a method for real-time human body reconstruction of a neural symbol distance field, including: Acquiring human body information of a target individual; constructing a gesture conditional neural network model according to human body information of a target individual; Training the attitude condition neural network model to obtain a target attitude condition neural network model after training; Inputting the query points into a trained target attitude condition neural network model, and outputting symbol distance field predicted values corresponding to the query points; The gesture conditional neural network model comprises a double encoder module, a feature fusion module and a FiLM conditional decoding module, wherein the double encoder module is used for extracting surface features and gesture features according to human body information of a target individual, the feature fusion module is used for carrying out feature fusion on the surface features and the gesture features to obtain condition vectors, and the FiLM conditional decoding module is used for outputting symbol distance field predicted values corresponding to query points according to the query points and the condition vectors. In a second aspect, an embodiment of the present invention provides an electronic device, including a memory and a processor, where the memory stores a computer program, and the processor implements the method for reconstructing a human body in real time from a neurosymbol distance field as in the first aspect or any one of the possible implementations of the first aspect when executing the computer program. The embodiment of the application provides a real-time human body reconstruction method of a nerve symbol distance field and electronic equipment, wherein the real-time human body reconstruction method of the nerve symbol distance field comprises the steps of obtaining human body information of a target individual, constructing a gesture conditional neural network model according to the human body information of the target individual, training the gesture conditional neural network model to obtain a trained target gesture conditional neural network model, inputting a query point into the trained target gesture conditional neural network model, and outputting a symbol distance field predicted value corresponding to the query point, the gesture conditional neural network model comprises a double encoder module, a feature fusion module and a FiLM condition decoding module, the double encoder module is used for extracting surface features and gesture features according to the human body information of the target individual, the feature fusion module is used for carrying out feature fusion on the surface features and the gesture featur