FPGA-Based CNN Acceleration on Zynq-7020 for Embedded Ship Recognition in Unmanned Surface Vehicles.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Embedded Systems Engineering

Background:

• Unmanned Surface Vehicles (USVs) require robust vision-based perception for navigation and surveillance.
• Onboard computing for USVs faces strict Size, Weight, and Power (SWaP) limitations.
• Deep Convolutional Neural Networks (CNNs) offer high recognition accuracy but demand significant computational resources, challenging embedded deployment.

Purpose of the Study:

• To develop and evaluate a hardware-software co-design architecture for accelerating CNNs on embedded platforms for USVs.
• To achieve energy-efficient, near-sensor processing for maritime perception tasks.
• To address the computational and memory constraints of deploying CNNs on low-cost embedded systems.

Main Methods:

• A heterogeneous ARM-FPGA system was utilized for CNN acceleration.
• A fully streaming hardware architecture based on line-buffered convolutions and AXI-Stream dataflow was implemented on the FPGA.
• The ARM processing system managed lightweight configuration, scheduling, and data movement.

Main Results:

• The proposed co-design strategy demonstrated a balanced trade-off between throughput, resource utilization, and power consumption.
• Experimental results were obtained using representative CNN models trained on a maritime ship dataset.
• The architecture was evaluated on a Zynq-7020 System-on-Chip.

Conclusions:

• The developed hardware-software co-design is suitable for energy-efficient onboard perception in USVs.
• The approach effectively addresses the challenges of deploying CNNs under tight embedded constraints.
• This work provides a practical building block for enhancing perception capabilities in autonomous maritime systems.