Real-Time Multi-Scale Face Detector on Embedded Devices.

Area of Science:

• Computer Vision
• Embedded Systems Engineering
• Machine Learning

Background:

• Real-time face detection on resource-limited embedded devices is a significant challenge.
• Existing face detectors often struggle to balance accuracy and speed on low-power hardware.

Purpose of the Study:

• To propose EagleEye, an efficient face detector for embedded devices.
• To achieve a favorable trade-off between high accuracy and fast detection speeds.
• To demonstrate superior performance on platforms like the Raspberry Pi 3b+.

Main Methods:

• Developed EagleEye with low floating-point operations per second (FLOPS) and sufficient network capacity.
• Implemented five key strategies: convolution factorization, successive downsampling convolutions, an efficient context module, an information-preserving activation function, and focal loss.
• Optimized for computational efficiency and accuracy improvement without significant FLOPS increase.

Main Results:

• EagleEye demonstrates high accuracy and fast runtime efficiency on embedded devices.
• Outperforms other face detectors with comparable computational costs.
• Successfully balances accuracy and speed on low-power hardware such as the Raspberry Pi 3b+.

Conclusions:

• EagleEye offers an effective solution for real-time face detection on computation-resource-limited embedded systems.
• The proposed strategies provide a blueprint for designing efficient and accurate lightweight deep learning models.
• This work advances the field of embedded computer vision by enabling robust face analysis in constrained environments.