A Lightweight FPGA-Based Hardware for Infrared Small-Target Real-Time Detection Using CNN

Abstract

This paper presents a resource-constrained, real-time on-board FPGA-based system architecture and design approach for unmanned aerial vehicle (UAV) computing hardware tailored for infrared small-target detection. Modern autonomous UAV systems require the high-throughput processing of video streams, demanding real-time performance that exceeds 150 frames per second (FPS). Concurrently, deploying convolutional neural networks (CNN) directly onto flight-ready hardware is heavily restricted by system-level performance, power, weight, volume, and cost (PPWVC) boundaries throughout the hardware design cycle. To overcome these rigid embedded system constraints, we integrate conventional digital signal processing techniques with lightweight convolutional neural network topologies into a unified, multi-stage processing pipeline. The end-to-end system design approach is demonstrated from initial parameter trade-offs and structural constraints to the final generation of a production-ready FPGA bitstream. This design approach demonstrates the rapid deployment of a high-speed, heterogeneous processing platform comprising an embedded CPU, programmable logic fabric, and a dedicated CNN accelerator. Open-access, single-frame infrared small-target datasets are utilised to validate the target detection performance and operational robustness of the proposed architecture in resource-constrained environments. Furthermore, a comparative analysis against commercial off-the-shelf computing platforms demonstrates that the proposed architecture satisfies strict PPWVC boundaries, delivering superior execution throughput alongside minimised power consumption.