Vehicle Object Detection Based On FPGA And Machine Vision

In recent years,major enterprises and university research institutions have begun research on autonomous driving and deep learning techniques.Vehicle target detection technology is a very important part of autonomous driving technology.It is necessary to consider not only the accuracy of recognition but also the algorithm.performance.Performance problems are also a big challenge in current research.There are two ways to improve the performance of the algorithm.One is to achieve software acceleration through algorithm optimization such as network optimization,parameter simplification,etc.,and the other is through the selection of operations.A powerful carrier enables hardware acceleration.Firstly,the paper studies all the algorithms of vehicle target detection on the CPU of the PC,analyzes the performance of each part,and explores the hardware acceleration of the time-consuming and accelerated feasible part.After the research,the convolutional neural network operation part is adopted.Hardware Acceleration.However,considering the autopilot scenario,the GPU algorithm accelerates the power consumption and cannot guarantee the energy saving and security of the vehicle.The flexibility of the ASIC algorithm is poor,and the timely replacement of the new algorithm cannot be guaranteed.Only FPGAs combine high flexibility,low power consumption and powerful computing power.This design studies the performance improvement of the neural network classification algorithm through the architecture of the CPU+SOC chip of the PC.The SOC uses the Xilinx high-performance chip ZYNQ7035,and the processor part is responsible for interacting with the PC through the high-speed Gigabit Ethernet.And data flow control,the hardware logic part is responsible for the hardware acceleration implementation of the neural network classification algorithm.This design has optimized the data flow control to make the data throughput nearly doubled.The focus of this research is to use FPGA to accelerate the algorithm of neural network,including color image data stream serial conversion and transmission,convolution layer,activation function layer,pooling layer and all-layer layer operation.In the case of parameters,when the impact on the accuracy and the recall rate is small,the fixed point is used to save the storage resources and computing resources of the weight parameter,the offset parameter and the operation result.The storage resource uses a BRAM resource for storing the trained weight parameters and the result of the convolution operation.The computing resources are DSP resources.The design occupies 800 DSP resources,and 800 pairs of weight parameters and data can be simultaneously multiplied.Programming uses data flow control,pipeline ideas,multi-module parallel,additive tree,cache and other parallel programming ideas to complete the algorithm acceleration;using fixed-point,resource-time multiplexing and other hardware programming ideas to complete the area optimization;using state machine single heat code Hardware programming ideas such as register segmentation complete timing optimization.The final implementation of the convolutional neural network of the FPGA infers that the acceleration performance is 3224 times that of the ARM CPU and 15 times that of the Intel core i7-8750 CPU.Therefore,the FPGA has better acceleration effect on the classification operation of the neural network;and its programming flexibility is high.In other deep learning scenarios,it is only necessary to update the weight parameters and increase the state machine to accelerate other neural network operations;It is less expensive and can be applied in vehicle detection scenarios.