CNN-based Traffic Sign Detection And Recognition Algorithm Design And FPGA Verification

With the development of the economy and society,the people's living standards have been greatly improved,and the vehicle have become daily necessities,which lead to serious traffic safety and traffic congestion problems.As a part of intelligent transportation system,vehicle traffic sign detection and recognition system has been paid attention by intelligent traffic researchers.Compared with traditional detection and recognition methods,CNN-based traffic sign detection and recognition algorithm has great advantages in scalability and robustness,but it has great optimization space in terms of accuracy,computation and storage space.Firstly,this thesis briefly introduces several typical algorithms in the field of traffic sign detection and recognition,and compares their advantages and disadvantages.The YOLOv2-tiny convolution neural network with high detection speed and suitable for hardware acceleration is selected as the basis of optimization.Aiming at the disadvantage of its poor detection effect on small targets,a traffic sign detection and recognition algorithm based on multi-scale feature fusion of convolution neural network is designed.In order to verify the effectiveness of the algorithm,two kinds of neural networks with different feature scales and YOLOv2-tiny are trained and analyzed based on Tsinghua-Tencent open traffic sign data set.The experimental results show that the detection and recognition effect of small targets has been significantly improved by multi-scale network.Finally,an algorithm verification system based on FPGA is designed.A high performance hardware acceleration scheme of convolution neural network is realized by using intraconvolution ganglion point parallelism,inter-convolution ganglion point parallelism and feature parameter reuse.Finally,data transceiver of each convolution layer is implemented in embedded hardware system,and the verification of traffic sign detection and recognition algorithm is realized.The convolution neural network designed and trained in this thesis achieves 92.62% recall rate and 90.30% accuracy rate on Tsinghua-Tencent open data set.Compared with YOLOv2-tiny convolution neural network,the recall rate and accuracy rate of small target detection are increased by 14.36% and 8.5% respectively.The designed convolutional neural network accelerator achieves 24.96 GOPS throughput and 11.79 GOPS/W power consumption ratio on Zynq-XC7Z020 FPGA,and achieves 3.4 frames per second detection speed.The algorithm and its hardware verification system designed in this thesis have the characteristics of good accuracy,small platform size and high resource utilization.It can provide a reliable reference scheme for the application of traffic sign detection and recognition system.