Implementation Of Target Detection System Based On Deep Learning On FPGA

As a branch of computer vision,target detection is a very hot research field at present,and its application prospect is very broad.It can be used in various fields such as intelligent transportation,intelligent monitoring,and autonomous driving.Its main task is to obtain the category information and location information of the target in the image data,and needs to solve the positioning problem and the recognition problem.The target detection method based on deep learning is a research hotspot in this field in recent years,and its detection accuracy and detection speed are greatly improved compared with traditional detection methods.However,since deep learning is a computationally intensive method,it has great requirements on hardware computing power.At present,commonly used computing graphics cards are used for testing.Due to the high price and high power consumption of high-performance computing servers,the cost of building them is high and it is not easy to expand.Therefore,how to expand the total amount of parallel processing and reduce the overall system construction cost through other methods has become an urgent problem to be solved.This paper designs a real-time target detection system based on an ARM + FPGA heterogeneous platform.The system can be used as an edge computing device to complete data processing on the near data side.To reduce the construction cost and computing pressure of the server.This article first elaborated the research achievements and research status of target detection and deep learning.After that,the specific model of the hardware platform used in the design was selected,and then the hardware support package based on the SDSOC kit was built.After the darknet framework was transplanted to the platform,it was successfully tested.In order to solve the problem of slow system running speed,a system optimization scheme was designed,and a general hardware convolutional layer accelerator was designed on the hardware side through the SDSOC development kit.At the algorithm end,combined with the actual needs of the system,retrain the yolo neural network.Finally,the actual effect of the system optimization scheme is tested.This design can achieve the same detection effect as the server side,and its power consumption is greatly reduced.In terms of computing speed,compared to the simple ARM-side calculation,it can achieve the effect of speeding up several times.