A Manycore Vision Processor for Real-Time Smart Cameras

Area of Science:

• Computer Engineering
• Embedded Systems
• Image Processing

Background:

• Real-time image processing is crucial for modern technologies like IoT, AR, and Industry 4.0.
• Commercial cameras struggle with the high data volumes and processing demands of these applications.
• Smart cameras require localized, efficient, and high-throughput image processing capabilities.

Purpose of the Study:

• To design and implement a manycore vision processor architecture for smart cameras.
• To address the limitations of commercial cameras in handling massive real-time image data.
• To create a flexible and performant hardware/software solution for advanced vision tasks.

Main Methods:

• Developed a manycore vision processor architecture featuring distributed processing elements and memories connected via a Network-on-Chip.
• Implemented the architecture as a Field-Programmable Gate Array (FPGA) overlay for optimized hardware utilization.
• Characterized the architecture's performance across various configurations (1 to 81 processing elements) and compared it with existing literature.

Main Results:

• The proposed architecture demonstrates efficient hardware utilization and high operating frequencies.
• Achieved significant processing frame rates, scalable with the number of processing elements.
• Validated the architecture's flexibility and efficiency using a System-on-Chip (SoC) integrating an FPGA and a general-purpose processor.

Conclusions:

• The manycore vision processor architecture successfully balances programmability and performance.
• It presents a suitable and advanced alternative for next-generation smart cameras.
• The design enables efficient real-time image processing for demanding cyber-physical and industrial applications.