SoC FPGA Accelerated Sub-Optimized Binary Fully Convolutional Neural Network for Robotic Floor Region Segmentation

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate segmentation of floor regions is crucial for robot navigation in complex indoor environments.
• Existing methods may struggle with real-time performance on embedded systems.
• Efficient visual perception is key for autonomous robot operation.

Purpose of the Study:

• To propose a novel Binary Fully Convolutional Neural Network (B-FCN) for precise robotic floor region segmentation.
• To optimize the B-FCN using the Taguchi method for enhanced accuracy.
• To accelerate the B-FCN for real-time computation on embedded platforms.

Main Methods:

• Development of a Binary Fully Convolutional Neural Network (B-FCN).
• Application of the Taguchi method for sub-optimization of the B-FCN architecture.
• Utilization of a PYNQ FPGA platform with heterogeneous computing for acceleration.
• Training the model with 6000 datasets for improved accuracy and convergence.

Main Results:

• Achieved an average segmentation accuracy of 84.80% for robotic floor regions.
• Demonstrated efficient reduction in BRAM size (0.5-1%) through FPGA synthesis.
• High GOPS/W (Giga Operations Per Second per Watt) achieved, indicating power efficiency.
• Enabled real-time computation suitable for embedded robotic vision.

Conclusions:

• The proposed B-FCN offers precise floor region segmentation for complex indoor environments.
• The accelerated architecture is ideal for low-power embedded devices, benefiting robot navigation, route planning, and motion planning.
• The methodology enhances robot vision capabilities, enabling better path searching and shortest path problem solutions.