Hardware Accelerator Design And Implementation Of Object Detection For Unmanned Aerial Vehicles

Nowadays,the technology of Unmanned Aerial Vehicle(UAV)is becoming more and more mature,and has made important breakthroughs in both military and civilian fields.Its typical application scenarios include logistics,agricultural plant protection,inspection security,emergency rescue,etc.With the development of AI(Artificial Intelligence)technology,UAVs are also widely used in target detection systems based on AI.An implementation of this system is that the UAV collects image information and transmits to a remote server.The server uses Convolutional Neural Network to achieve specific target recognition,and guides subsequent operations.In this situation,there are two disadvantages:First,the mass of data transmission leads to insufficient bandwidth and even data loss.Second,the data transmission delay is long,which affects the timely execution of instructions.In order to solve these problems,some scholars propose to deploy a neural network on UAV,thereby reducing the amount of data transmission and delay.But this solution faces the challenge of implementing a high-load neural network in a lightweight UAV system-on-chipIn order to implement a neural network object detection system on the UAV,this paper selects a high-performance and low-power PYNQ(ARM+FPGA)as a computing platform to deploy a neural network,and implements some optimization strategies to increase speed of specific object detection accuracy.Our designed is mainly divided into three parts,namely image preprocessing on ARM,hardware acceleration with FPGA,and result integration optimization on ARM First,in order to make the input form of the data beneficial to the processing of the neural network in FPGA,the original image is cut and transformed to improve the network operation efficiency.Next in FPGA part,considering the limited internal computing and storage resources,this paper converts the proposed Convolutional Neural Network(CNN)into a Binary Neural Network(BNN)to complete object detection and increase FPS.Besides,this paper uses HLS optimization strategies and AXI4 protocol to improve FPGA data throughput.Finally,the FPGA returns the calculation results to the ARM,using the position correction strategy based on breadth first search to optimize the target position,and obtains the final result.This paper explores the combination of UAV and neural network,which has good social significance and application value.The design can also be used for marine object recognition,automatic driving,industrial specific object detection and other scenarios.