摘要
针对粒子滤波网络(PF-Net)模型用于高度不确定环境中进行机器人视觉定位需要大量粒子才能精准定位的问题,提出一种自适应软重采样粒子滤波网络(ASRPF-Net)。为更好提取观测信息和地图信息,采用卷积神经网络(CNN)学习观测图像信息和地图信息,能够有效降低学习的复杂性,提高粒子权重的准确性。在重采样过程中加入决策,提出一种自适应软重采样方法,使模型能够判断是否需要进行重采样,达到减少粒子数量,减缓粒子集退化,提高机器人定位准确性的目的。在House3D和DeepMind Lab仿真环境上的实验结果表明,该方法比其它网络模型具有更高的定位精度和更好的鲁棒性。
To solve the problem that the particle filter network(PF-Net)model for robot visual localization needs a large number of particles to accurately localize in highly uncertain environment,an adaptive soft-resampling particle filter network(ASRPF-Net)was proposed.To better extract observation information and map information,convolutional neural network(CNN)was used to learn observation image information and map information,which effectively reduced the complexity of learning and improved the accuracy of particle weights.The decision was added in the resampling process,and an adaptive soft-resampling method was proposed to enable the model to determine whether resampling was required,so as to reduce the number of particles,slow down the degradation of the particle set,and improve the accuracy of robot localization.Experimental results on House3D and DeepMind Lab simulation environments show that the proposed method has higher localization accuracy and better robustness than other network models.
作者
刘艳丽
尹慧君
张恒
LIU Yan-li;YIN Hui-jun;ZHANG Heng(School of Information Engineering,East China Jiaotong University,Nanchang 330013,China;School of Electronic Information Engineering,Shanghai Dianji University,Shanghai 201306,China)
出处
《计算机工程与设计》
北大核心
2022年第12期3503-3512,共10页
Computer Engineering and Design
基金
国家自然科学基金项目(61963017、61663010)
江西省科技创新杰出青年人才基金项目(20192BCBL23004)。
关键词
机器人定位
粒子滤波
深度学习
循环神经网络
卷积神经网络
重采样
视觉定位
robot localization
particle filter
deep learning
recurrent neural network
convolutional neural network
resampling
visual localization
