Research And Implementation Of Dynamic Reconfigurable Hardware Acceleration System For Multi-Object Detection Algorithm

With the development of artificial intelligence technology in the information age and the tense international situation at present,the defense and military of various countries in the world are accelerating the promotion of new combat modes of intelligence and unmanned aerial vehicles,among which the demand for target detection is more urgent.In the era of artificial intelligence preparedness,effective detection of hostile military targets is of great significance for solving national defense problems,improving our national security and enhancing military combat capabilities.To achieve automatic target recognition on mobile devices in various environments,resear-chers have focused on deploying target detection algorithms on embedded edge computing platforms.Although target detection algorithm accuracy and efficiency have improved,the increasing number of network layers and parameter sizes make deploying algorithms on edge computing platforms challenging and lengthen detection times for single images.To address the challenges of deploying military aircraft target detection algorithms on edge hardware devices and the difficulty of real-time detection,this thesis studies a dynamic hardware acceleration system for multi-object detection algorithms.Firstly,the thesis designs a lightweight YOLOv5s model and investigates a new lightweight target detection model,GSConv-YOLOv5s,using the SIoU regression loss function and the Hardswish activation function to reduce model parameters and improve model performance with only a slight loss in detection accuracy.Secondly,a dynamic hardware acceleration system architecture for multi-object detection algorithm hardware,called EDRCA-YOLO,is studied.This includes a dynamic hardware subsystem for image preprocessing algorithm acceleration and a DPU-based hardware acceleration subsystem for deploying the GSConv-YOLOv5s inference network.Then,the thesis studies a target detection network inference accelerator solution based on Vitis AI.This solution compresses the parameter size of the target detection model by freezing,calibrating,quantizing,and compiling parameters,and then generates executable files to accelerate network inference operations by calling application programs in the PetaLinux operating environment deployed on the PS side.Finally,the EDRCA-YOLO system architecture is deployed on the KV260 hardware platform,and Python and C++application programs are used on the PS side to schedule the two subsystems in the PL side to achieve image preprocessing algorithm acceleration and GSConvYOLOv5s inference network operation acceleration.The thesis trains the GSConv-YOLOv5s network on a Tesla T4 GPU with a self-made military aircraft target dataset FTDS,and deploys the Vitis AI IDE on Ubuntu 18.04.The Vitis AI tool library is then used in the PyTorch deep learning framework to freezing,calibration,quantization and compilation.The required IP cores for the image enhancement algorithms were also generated at Vitis HLS and the configurable profiles of the operators were generated on the Vivado development platform using dynamic reconfiguration technology.The PYNQ-Kria hardware-software co-development platform was built on XILINX KV260 FPGA hardware to deploy the target detection algorithm inference network and the image pre-processing algorithm by scheduling the DPU IP cores on the PL side and loading the image enhancement operator profiles through the application on the PS side.According to the implementation results,the dynamically reconfigured hardware acceleration system for multi-target detection algorithms studied in the thesis achieves image enhancement and enhanced visualization of detection images from different environments in runtime,and also obtains an average accuracy of 46.6%and detection rate of 30.0 FPS,with an average accuracy loss of 2.7%compared to pre-deployment,while the system operates with a power consumption of only 6.035W.The system provides a viable solution for the deployment of military aircraft target detection systems on edge-mobile devices in a variable battlefield environment.