无人车地面目标识别及其优化技术研究

Research on Ground Target Recognition and Optimization Technology of Unmanned Vehicle

针对智能目标识别算法在无人车嵌入式硬件上的应用需求,研究基于卷积神经网络的地面目标识别算法及其硬件适应性优化技术,提出了基于双正则项的自适应网络裁剪优化算法及面向FPGA的神经网络INT8量化优化算法。针对裁剪及量化后的YOLO V3算法,完成了基于Xilinx公司的UltraScale+MPSoC系列的XCZU7EV器件验证平台的智能算法程序部署,并利用无人车的可见光数据集和红外数据集进行了仿真验证。结果表明,YOLO V3算法在两类优化算法有效结合并保持网络精度的前提下,在无人车嵌入式硬件平台上可获得4.5倍的加速比,能够满足无人车地面目标识别的精度与实时性要求。所提出的优化算法思想为神经网络在嵌入式硬件平台上的部署提供了新的技术思路。

Aiming at the application requirements of intelligent target recognition algorithm in unmanned vehicle,the ground target recognition algorithm based on CNN(Convolutional Neural Networks)and its optimization technology are studied.An adaptive network pruning optimization algorithm based on double regularization terms,and an INT8 quantization optimization algorithm for FPGA are proposed.Aiming at the pruned and INT8 quantified YOLO V3 algorithm,the intelligent algorithm program deployment of XCZU7EV device verification platform based on Xilinx’s Ultrascale+MPSoC series is completed,and the simulation verification is carried out by using the visible light data set and infrared data set of unmanned vehicle.The experimental results show that with the effective combination of the two kinds of optimization algorithms proposed in this paper,YOLO V3 algorithm can achieve 4.5 times acceleration ratio on the FPGA platform while maintaining the network accuracy,which can meet the accuracy and real-time requirements of ground target recognition of unmanned vehicle.The optimization algorithm proposed in this paper provides a new technical idea for the deployment of neural network on embedded hardware platform.