Feature Point Detection Using 3D Model Projection and Error Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for detecting feature points in images, such as ASM and AAM, face challenges with high computational complexity and low robustness, especially when dealing with varying face orientations and poor lighting conditions, which hinders accurate and efficient feature point detection for applications like driver monitoring systems.
Innovation Solution
A feature point detecting device that uses a three-dimensional model to estimate the position of feature points by projecting nodes onto an image plane, acquiring feature values through non-homogeneous and low-density sampling, and employing canonical correlation analysis for error estimation, eliminating the need for shape correcting processes and high-density sampling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional methods (ASM, AAM) are used for feature point detection, then detection accuracy can be maintained under ideal conditions, but computational complexity increases and robustness decreases under varying face orientations and poor lighting conditions
Solution Approach 1:
The patent segments the face detection process into distinct functional modules: a face shape model creation unit that pre-processes training images to establish geometric constraints, a feature point detection unit that applies template matching with geometric constraints, and a face orientation estimation unit. This segmentation allows each module to specialize in specific tasks, improving overall robustness while managing computational complexity through distributed processing.
Solution Approach 2:
The patent applies preliminary action by pre-processing training images to create a face shape model that encodes geometric constraints before actual detection occurs. The model creation unit pre-computes relationships between feature points from multiple training images, storing this knowledge in a reusable format. This preliminary preparation enables the detection phase to operate more efficiently and robustly without repeating computationally intensive operations.
2Measurement precision
If high-density sampling is performed to maintain detection accuracy, then measurement precision improves, but computational time increases significantly
Solution Approach 1:
The patent applies local quality by implementing non-uniform sampling density across different regions of the face. The sampling density is adjusted locally based on the importance and variability of feature points in different facial regions. Critical regions receive higher sampling density while less critical regions use lower density, optimizing the balance between detection accuracy and computational efficiency.
Solution Approach 2:
The patent changes the sampling parameter from uniform high-density to non-uniform low-density sampling. By modifying the sampling strategy to use fewer points distributed non-uniformly across the face, the system reduces computational time while maintaining detection accuracy through the compensating effect of geometric constraints from the face shape model.
3Productivity
If template matching is used with geometric constraints to reduce search points, then productivity improves, but detection precision decreases when feature points are hidden or lighting conditions differ significantly
Solution Approach 1:
The patent applies dynamics by making the template matching process adaptive rather than static. The geometric constraints from the face shape model dynamically guide the template matching, allowing the search to focus on plausible regions while maintaining flexibility to handle variations in lighting and occlusion. This dynamic approach improves both speed and accuracy compared to static template matching.
Data Source
Figure 1A~1B
Figure 2
Figure 3
AI summary
The positions of the feature points are detected at high speed from an image. Information on the correlation of a difference between a correct model in which each node is arranged at a correct position of a feature point and an error model in which one of the nodes is arranged at a wrong position, and a node feature value acquired based on the error model is acquired in advance using a learning image. In detecting the feature point from the input image, the three dimensional model in which the three dimensional positions of a plurality of nodes are defined is created, each node is projected onto an input image, a node feature value is acquired from a projection point, and an error estimated amount indicating the displacement of the position of each current node and the position of the corresponding feature point is acquired based on the node feature value and the information of the learnt correlation, and the three dimensional position of each feature point in the input image is estimated based on the error estimated amount and the position of each current node.