基于改进SegNet网络的障碍物检测算法研究
OBSTACLE DETECTION ALGORITHM BASED ON IMPROVED SEGNET NETWORK
胡均平 1张洪伟 1黄强 1向思平1
作者信息
- 1. 中南大学机电工程学院 湖南长沙 410006
- 折叠
摘要
为了提高自动驾驶车辆的多类障碍物检测的精度,提出一种改进的SegNet神经网络算法.该算法在SegNet网络的基础上结合残差网络和多尺度融合算法,提高分类的准确性.对网络训练策略进行改进,采用 自设置对比度归一化算法、学习率调整算法和类平衡算法提高网络的鲁棒性和收敛速度.通过在不同场景下进行实验,结果表明,相较于SegNet神经网络,改进后的SegNet神经网络的像素平均精度(PPA)从85％提高至97％;平均交并比(MIOU)从76％提升至90％.
Abstract
In order to improve the accuracy of multi-class obstacle detection for autonomous vehicles,an improved SegNet neural network algorithm is proposed.Based on SegNet,the algorithm combined residual network and multi-scale fusion algorithm to improve the accuracy of classification.The self-setting contrast normalization algorithm,learning rate adjustment algorithm and class balance algorithm were used to improve the robustness and convergence speed of the network.Through experiments in different scenes,the results show that compared with SegNet neural network,the per pixel accuracy of the improved SegNet neural network is improved from 85％to 97％,and the mean intersection over union is improved from 76％to 90％.
关键词
自动驾驶/障碍物检测/SegNet网络/残差网络/多尺度融合算法Key words
Autonomous driving/Obstacle detection/SegNet network/Residual network/Multiscale fusion algorithm引用本文复制引用
基金项目
国家自然科学基金(51175518)
湖南省科技型中小企业技术创新基金(12C26214305029)
出版年
2024
国家科技期刊平台
NSTL
万方数据