Real-Time Detection of Cook Assistant Overalls Based on Embedded Reasoning.

Area of Science:

• Computer Vision
• Embedded Systems
• Deep Learning

Background:

• Convolutional neural networks (CNNs) are vital for image and speech recognition but often too complex for resource-constrained embedded devices.
• High parameter counts and computational demands limit deep learning applications in low-power, sensitive environments.
• Existing CNN models face challenges in real-time performance and deployment on edge devices.

Purpose of the Study:

• To propose a real-time detection scheme for cook assistant overalls using the Hi3559A embedded processor.
• To optimize a YOLOv3 network for efficient execution on embedded hardware.
• To enable local, real-time image detection on devices with limited computational power and strict power constraints.

Main Methods:

• Utilized YOLOv3 as a benchmark network and adapted it for the Hi3559A embedded processor.
• Employed network model optimization techniques, including purposeful cropping and segmentation.
• Leveraged the processor's hardware acceleration resources and parallel processing capabilities.
• Focused on in-depth optimization tailored to the specific processor architecture.

Main Results:

• Achieved a recognition speed of approximately 28 frames per second on the embedded end, exceeding the design requirement of 25 FPS.
• Demonstrated accurate image recognition of cook assistant overalls.
• Operated effectively within a low power consumption environment of 5.5 W.

Conclusions:

• The optimized YOLOv3 network effectively performs real-time detection of cook assistant overalls on the Hi3559A embedded processor.
• Network optimization and hardware acceleration successfully addressed the limitations of deploying deep learning on embedded systems.
• The proposed scheme enables practical application of deep learning for object identification in resource-limited environments like back kitchens.