Reproduce “Real Time FPGA Based CNNs for Detection, Classification, and Tracking in Autonomous Systems: State of the Art Designs and Optimizations” on OpenResearch

About this paper

Safa Mohammed Sali, Mahmoud Meribout, Ashiyana Abdul Majeed

Khalifa University researchers provide a comprehensive review of real-time FPGA-based Convolutional Neural Networks for detection, classification, and tracking in autonomous systems, analyzing state-of-the-art designs, hardware architectures, optimization techniques, and software tools. The work synthesizes current advancements to guide the development of power-efficient and high-performance vision systems for embedded and edge applications.

Key insights

• FPGAs offer a balanced alternative to GPUs and ASICs for edge AI, providing reconfigurability, deterministic latency, and lower power consumption suitable for low-batch, streaming, and latency-sensitive applications.
• Efficient CNN deployment on FPGAs hinges on exploiting various forms of parallelism (kernel-level, spatial, channel-level), spatial dataflow architectures (e.g., systolic arrays), and advanced pipelining and layer fusion techniques.
• Precision scaling through quantization (e.g., INT8, INT4) and hardware-aware pruning and sparsity methods are crucial for reducing memory bandwidth and resource utilization while maintaining acceptable accuracy on FPGA platforms.

Summary by alphaXiv.