Implementation of Real-Time Object Detection on FPGA

Abstract

Object detection is a common and challenging problem in today's world. Also due to the rapid development of deep learning employing underlying deep models over the past ten years, many researchers have investigated and contributed to improving the performance of object identification and associated tasks such as object classification, localization, and segmentation. Inference time and detection accuracy are used to rate the effectiveness of any object detector. In terms of detection accuracy, it is observed that the two-stage detectors outperform the single-stage detector however, single-stage detectors outperform their rivals in terms of inference time. Additionally, You Only Look Once (YOLO) and its versions have improved detection accuracy, some-times surpassing two-stage detectors. In this study, we performed a co-design methodology for hardware and software (HW/SW) aimed at Central Processing Unit (CPU) + Field Programmable Gate Array (FPGA) based heterogeneous platform which is Xilinx FPGA Kria KV260. A CNN-based algorithm is first extended to the YOLOv4, v5L, and v5s framework before being implemented on the Kria KV260, an out-of-the-box platform for developing advanced vision applications. A reduced computation time is observed for the deployment of the Machine Learning model on Kria KV260.