基于类脑认知与贝叶斯估计的仿生建图方法

Bionic mapping method based on brain-inspired cognition and Bayesian estimation

针对同步定位与地图构建(SLAM)过程中因传感器累积误差、干扰等情况造成定位精度不高的问题,借鉴哺乳动物海马空间对多源信息的认知与整合机理,提出一种新型空间认知模型和融合贝叶斯估计的误差修正方法.首先,建立视觉线索的模拟机制,并对头朝向细胞和3D网格细胞进行建模;其次,构建反赫布学习递归神经网络完成对特定位置的编码,实现机器人对时空经验信息稳定的表达;最后,采用贝叶斯估计原理,建立原位置细胞邻域空间,提出结合位置细胞放电率的误差修正策略.与NeuroSLAM,OKVIS和VINS算法在EuRoC公开数据集上进行对比测试,结果表明:本文算法能够整合多源环境信息实现认知地图构建,最大绝对平移误差为0.727 m,最大绝对旋转误差为0.442.

Aiming at the problem of low positioning accuracy caused by sensor cumulative error,interference and other emergencies in the process of simultaneous localization and mapping(SLAM),inspired by multi-source information integrated and cognitive mechanism of hippocampal spatial in mammals,a new spatial cognitive model and an error correction method integrated with Bayesian estimation was proposed.First,the simulation mechanism of visual cues was established,and the head-direction cells and 3D grid cells were modeled.Then,the anti Hebbian learning recurrent neural network was constructed to complete the coding of specific positions,and realize the stable expression of spatiotemporal experience information of the robot.Finally,by using Bayesian estimation principle,the neighborhood space of the original place cells was established,and an error correction strategy combined with the firing rate of the place cells was proposed.Compared with NeuroSLAM,OKVIS and VINS on the EuRoc public dataset,results show that proposed algorithm can integrate multi-source environmental information to construct cognitive map,while the maximum absolute translation error is 0.727 m,and the maximum absolute rotation error is 0.442.