一种融合DGSOM神经网络的仿生算法研究

PDF下载HTML

Title:

A bio-inspired algorithm integrated with DGSOM neural network

作者:

许曈, 凌有铸, 陈孟元

安徽工程大学 安徽省电气传动与控制重点实验室, 安徽 芜湖 241000

Author(s):

XU Tong, LING Youzhu, CHEN Mengyuan

Anhui Key Laboratory of Electric Drive and Control, Anhui Polytechnic University, Wuhu 241000, China

关键词:

RatSLAM模型; DGSOM神经网络; 同步定位与地图构建; 闭环检测; 准确率; 召回率

Keywords:

RatSLAM model; DGSOM neural network; simultaneous localization and mapping; closed-loop detection; precision rate; recall rate

分类号:

TP242.6;TP183

DOI:

10.11992/tis.201704038

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RatSLAM模型中。实验表明，提出的DGSOM-RatSLAM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-RatSLAM模型的准确率、召回率及F₁值分别为94.74%、86.88%和90.64%，高斯噪声干扰下Gauss-DGSOM-RatSLAM模型的准确率、召回率及F₁值分别为86.70%、80.25%、83.35%。

Abstract:

Based on physiology and brain science, self-organizing-map (SOM) neural networks can learn and autonomously draw topological maps, but the initial SOM network structure must be repeatedly tested, so the real-time characteristics of the system cannot be assured. In this paper, we built a dynamic growing self-organizing map (DGSOM) based on direction and feature parameters that reduces network training times by the introduction of the direction parameter and decreases system complexity and avoids perceptual aliasing by the introduction of the feature parameter. By introducing the feature parameter, we can avoid perception confusion. We applied the proposed model to the view cells of the simultaneous localization and mapping system (SLAM) known as RatSLAM, proposed by Milford et al. Our experimental results show that the proposed DGSOM-RatSLAM model can decrease the complexity of the system by reducing the quantity of view cells and realize closed-loop detection earlier by matching the scene with view cells and detecting on the activity of the pose cells. We found the precision rate, recall rate, and F₁ value of the DGSOM-RatSLAM model to reach 94.74%, 86.88%, and 90.64%, respectively, and those of the Gauss-DGSOM-RatSLAM model to reach 86.70%, 80.25%, and 83.35%, respectively.