FPGA-based prototyping of drone detection algorithm

Abstract

In recent years, deep learning, especially convolutional neural network, has received much attention as a method of target detection. And better implementation of neural networks using FPGA is also one of the main research directions. Moreover, the detection of small targets, such as drones, is being applied in a wider range of industrial scenarios. However, the complexity of different network models and the limitation of resources on FPGA are also limiting factors of the improvement works. In this thesis, the simulation of detection of a dataset containing drone images is implemented on FPGA by using a simplified CNN network structure, while optimizing for the limited storage and computational resources of FPGA and utilizing parallelized pipeline processing. In addition, future trends for similar research are discussed with respect to the shortcomings of the design in this thesis. A performance of approximately 270ms for processing one image is achieved with an accuracy of approximately 96.4% on simulation platform, driven by one 200MHz clock.