Real-Time Ecological Surveillance: An FPGA-Based Object Detection Accelerator For Sustainable Edge Devices

Abstract

This research work proposed the design and implementation of a dedicated object detection FPGA-based accelerator for edge devices with limited computational resources. The system is built with a Xilinx Zynq FPGA-ARM SoC, using the extensive parallelism in the Programmable Logic (PL) to accelerate deep learning computations, and leveraging the ARM Processing Systems (ARM-PS) for control operations. Object detection backend: Tiny-YOLO for Linux, a quantised and pruned version of object detection algorithm optimized for FPGA execution. Due to this accelerator, the inference time is reduced drastically as it performs convolutional, activation and pooling operations on the PL. It captures video input from a USB web camera and overlay detection results on the video and display it in real time. Hence, the proposed work, examined the that implementation significantly outperforms a traditional CPU-based system, both in terms of throughput and energy efficiency and thus demonstrates the viability of using FPGA accelerators to process real-time edge devices.