Facial Feature Point Tracking Using Frame-to-Frame Error Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing facial feature point tracking methods fail to accurately represent real facial features due to limitations in tracking facial feature points across consecutive image frames using the same reference points.

Innovation Solution

A method that involves obtaining facial feature points in a previous image frame, calculating errors based on a preset error model and pixel differences, and using these errors to estimate facial feature points in subsequent frames, utilizing a random forest data structure for classification and regression.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If the same reference facial feature points are used to obtain facial feature points in each image frame, then the tracking process is simplified, but the tracking accuracy deteriorates because facial feature points continuously change across consecutive frames

Engineering Contradiction:
Improvetracking process complexityVSAvoidfacial feature point tracking accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent transforms the static reference point approach into a dynamic tracking system where the error model is continuously updated based on temporal differences between consecutive frames. The system adapts to changing facial features by incorporating motion information and temporal variations, allowing the tracking algorithm to follow dynamic facial expressions and movements while maintaining computational efficiency through the learned error patterns.

Inventive Principle:
Principle #15Dynamics

2Productivity

If facial feature points are obtained independently for each image frame using the same reference points, then the processing of each frame is independent and simple, but the continuity and accuracy of tracking across frames deteriorates

Engineering Contradiction:
Improveframe processing efficiencyVSAvoidtracking consistency across frames
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent implements a feedback mechanism where the error model learns from temporal differences between consecutive frames and uses this learned information to improve subsequent tracking. The system incorporates feedback from previous frame predictions and actual observations to continuously refine the error estimates, creating a closed-loop tracking system that maintains reliability across frames while preserving processing efficiency through the learned error patterns.

Inventive Principle:
Principle #23Feedback

3Measurement precision

If a complex tracking model that adapts to each frame's unique features is used, then tracking accuracy improves, but the computational complexity and processing time increase significantly

Engineering Contradiction:
Improvefacial feature point tracking accuracyVSAvoidtracking model complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent changes the approach from complex per-frame adaptation to a more efficient error-based parameter adjustment. Instead of re-running complex detection algorithms on each frame, the system modifies the tracking parameters by adding learned error corrections to predictions. This parameter change strategy maintains high accuracy by adapting to frame-specific variations while keeping computational complexity low through efficient error model updates and predictions.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3644219B1Human face feature point tracking method, device, storage medium and apparatus
Publication Date: 2026.04.15 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3644219B1 patent drawingFigure 1
  • EP3644219B1 patent drawingFigure 2
  • EP3644219B1 patent drawingFigure 3(a)~3(c)

AI summary

The present application discloses a human face feature point tracking method, device, storage medium and apparatus related to the technical field of image recognition. The method comprises: acquiring, in an image preceding an image undergoing tracking, a human face feature point; acquiring a human face feature point deviation on the basis of a predetermined deviation model and a pixel point in the image undergoing tracking, wherein the human face feature point deviation refers to a difference between first coordinates and second coordinates, the first coordinates are coordinates of the human face feature point in the image undergoing tracking, the second coordinates are coordinates of the human face feature point at a corresponding position in the preceding image, and the predetermined deviation model is used to indicate a relationship between a pixel point in a later image of adjacent images and the human face feature point deviation; and acquiring, on the basis of the human face feature point of the preceding image and the human face feature point deviation, a human face feature point of the image undergoing tracking. The present application uses a human face feature point of a preceding image as a reference, and thus is capable of more accurately estimating a human face feature point of an image undergoing tracking.