Face Tracking Using Subsampled Integral Images for Embedded Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face tracking systems in digital image acquisition devices face challenges in achieving real-time performance due to high computational requirements, leading to high false positive rates and inability to operate efficiently in resource-restricted embedded systems like handheld cameras, as they struggle to cover all scales and maintain accurate tracking across frames.
Innovation Solution
The method involves determining the initial and subsequent locations and sizes of faces in a preview image stream, predicting the face region in subsequent images, and adjusting acquisition parameters like white balance, focus, and exposure based on analyzed characteristics, using subsampled integral images and varying resolution to reduce processing overhead and improve tracking efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional face detection algorithms are used to cover all scales, then detection accuracy is improved, but computational complexity increases and real-time performance deteriorates
Solution Approach 1:
The patent divides the face detection process into two distinct stages: a detection stage that covers all scales to find candidate faces, and a tracking stage that focuses on confirmed faces across frames. This segmentation allows each stage to be optimized independently, reducing overall computational complexity while maintaining detection accuracy.
Solution Approach 2:
The patent performs preliminary face detection in preview frames to establish initial face locations and scales before the actual image capture. By pre-identifying face regions and their characteristics in advance, the system reduces the computational burden during the main acquisition phase while ensuring accurate face coverage.
2Measurement precision
If comprehensive face detection is performed in every frame, then tracking accuracy is improved, but processing time increases and real-time performance decreases
Solution Approach 1:
The patent maintains continuous face tracking across preview frames by carrying forward face location information from one frame to the next. Instead of performing complete detection in every frame, the system continuously updates face positions based on previous frame data, ensuring uninterrupted tracking while reducing redundant computations.
Solution Approach 2:
The system uses feedback from detected faces in preview frames to guide subsequent detection and tracking operations. Face locations and characteristics identified in early frames inform the search parameters and regions in later frames, creating a feedback loop that improves tracking accuracy while minimizing processing time through intelligent search optimization.
3Measurement precision
If high-resolution images are processed for face detection, then detection quality is improved, but computational load increases and efficiency decreases
Solution Approach 1:
The patent applies different processing qualities to different regions and stages: lower resolution processing for initial face detection in preview frames, and higher resolution processing only for confirmed face regions in the final capture. This local quality approach ensures detection quality is maintained where needed while improving overall processing efficiency.
Solution Approach 2:
The system performs partial face detection at lower resolution in preview frames, focusing computational resources on identifying face regions rather than analyzing every detail. This partial action approach provides sufficient detection quality for tracking purposes while significantly reducing the computational load compared to full-resolution processing of entire frames.
Data Source
AI summary
A method of tracking faces in an image stream with a digital image acquisition device includes receiving images from an image stream including faces, calculating corresponding integral images, and applying different subsets of face detection rectangles to the integral images to provide sets of candidate regions. The different subsets include candidate face regions of different sizes and/or locations within the images. The different candidate face regions from different images of the image stream are each tracked.