YOLOv3 Tiny-Based Human Detection and Tracking using Vitis AI and VVAS on FPGA

Real-time human detection systems have become increasingly popular due to advancements in science and technology. Numerous studies have been conducted to enhance object recognition capabilities with higher accuracy and faster processing speeds. However, current methods still face several limitations, particularly in balancing the accuracy and processing speed of recognition algorithms. One of the current approaches is to deploy Convolutional Neural Networks (CNN) on Field-Programmable Gate Arrays (FPGA) to improve performance. Previous studies have achieved certain results in using FPGAs, but they often require deep knowledge of hardware design, which presents challenges for practical implementation. To address this issue, this paper introduces the use of VVAS (Vitis Video Analytics SDK) and Vitis AI to implement CNNs on FPGA without the need for deep intervention in hardware design. We employ the YOLOv3-Tiny (You Only Look Once) model to detect humans in real-time on the Kria KV260 Kit. Experimental results show that the system can process and recognize objects with high accuracy while ensuring fast processing speeds for applications in real-world surveillance environments.