一种融合DGSOM神经网络的仿生算法研究

发布时间：2018-04-17 09:31

本文选题：RatSLAM模型 + DGSOM神经网络　；参考：《智能系统学报》2017年03期

【摘要】：基于生理学和脑科学研究成果提出的SOM神经网络仿生优化方法能够通过学习自主绘制出拓扑地图,但需通过大量的尝试确定其初始网络结构,无法保证系统的实时性。提出一种方向信息和特征信息构建的动态增长自组织特征网DGSOM,通过引入方向参数减少网络的训练次数,降低了系统复杂度,通过引入特征参数避免了感知混淆,并将该神经网络模型应用于澳大利亚Milford等提出的Rat SLAM模型中。实验表明,提出的DGSOM-Rat SLAM模型通过减少视觉细胞的数量降低系统的复杂度;通过视觉细胞的场景匹配实验和位姿细胞的活性状态实验证明该模型能够更快地实现闭环检测,提出的DGSOM-Rat SLAM模型的准确率、召回率及F_1值分别为94.74%、86.88%和90.64%,高斯噪声干扰下Gauss-DGSOM-Rat SLAM模型的准确率、召回率及F_1值分别为86.70%、80.25%、83.35%。
[Abstract]:The bionic optimization method of SOM neural network based on physiology and brain science can draw topological map by learning independently, but it needs a lot of attempts to determine its initial network structure, which can not guarantee the real-time performance of the system.A dynamic growth self-organizing feature net (DGSOM) based on direction information and feature information is proposed. By introducing directional parameters to reduce the training times of the network, the complexity of the system is reduced, and the perception confusion is avoided by introducing the feature parameters.The neural network model is applied to the Rat SLAM model proposed by Milford et al.The experiments show that the proposed DGSOM-Rat SLAM model can reduce the complexity of the system by reducing the number of visual cells, and the experiments of scene matching of visual cells and the active state of pose cells prove that the model can realize closed-loop detection more quickly.The accuracy, recall rate and FSC1 value of the DGSOM-Rat SLAM model are 94.744.88% and 90.64%, respectively. The accuracy of the Gauss-DGSOM-Rat SLAM model with Gao Si noise is 86.70% and 80.25%, respectively.
【作者单位】：安徽工程大学安徽省电气传动与控制重点实验室;
【基金】：安徽高校自然科学研究项目(KJ2016A794)
【分类号】：TP183;TP242

【相似文献】