| Simultaneous Location And Mapping(SLAM)means that a mobile robot in an unknown environment collects physical information about the current environment through various sensors it carries,models and maps the environment during its movement,and locates its position on the map at the same time.However,at the present stage,SLAM algorithms still have problems such as large computational volume and the tendency of mapping failure in complex environments.The level of intelligence of SLAM algorithms is still not as good as the animal’s ability to navigate.To this end,the development and research of new SLAM algorithms for mobile robots,especially the construction of bionic SLAM models drawing on the cognitive mechanism of the rat brain hippocampus,has been a research direction and focus in recent years.Australian scholars Milford et al.proposed a spatial navigation method based on rodent hippocampal structure,however,the method still has some defects,such as low accuracy of closed-loop detection under light visual transformation,large cumulative error under long operation,and interference from sudden dynamic obstacles also seriously affect the navigation effect of this navigation method,resulting in large deviation of the trajectory.To address the problems that the spatial navigation method based on rodent hippocampal structure is easily disturbed by complex environmental factors and the low accuracy of closed-loop detection under light visual transformation leads to large trajectory errors,this article improves the accuracy of closed-loop detection under sudden light transformation scenes by introducing color depth maps for loopback detection.Inspired by the mechanism of spatial cognition in the hippocampus of the rat brain,this article proposes a bionic SLAM algorithm based on the interest propensity mechanism.The algorithm adopts Lateral Anti-Hebbian Network(LAHN)to model the grid cells,and corrects the constructed grid cells by irregular environmental boundary information to improve the algorithm localization accuracy.The interest tendency mechanism is used to assign interest to the extracted significant regions to reduce the influence of redundant significant regions and improve the positioning accuracy of the system.The visual information extracted by the environment-aware model is fused with the place cells competed by the location-aware model to build a cognitive map with topological relations.In order to overcome the drawback that traditional SLAM algorithms are difficult to adapt to non-fixed data sets and not suitable for real-time online learning,this article introduces a visual bag-of-words model DGPBoVW(Dynamic growing and pruning Bag of Visual Word)for building maps of unknown environments based on a cognitive map construction algorithm imitating the rat brain hippocampus.The model can dynamically adjust the number of words in the visual bag of words in real time according to the complexity of the scene.Moreover,in order to prevent the camera from acquiring a large amount of redundant image information with too much similarity when the mobile robot stops or runs slowly,this article sets the distance threshold and angle threshold to select the preselected keyframes,and then uses the Tenengrad function to score and filter the pre-selected keyframes.Finally,the keyframes that satisfy the conditions are used as the input of the visual bag of words,and the similarity metric is applied to the historical keyframes.When a familiar scene is detected,the system occurs to close the loop then to correct the current position information of the robot.Finally,the algorithm of this article is validated in public KITTI,TUM datasets and real environments,and the proposed method is compared with the mainstream ORB-SLAM2 algorithm and the traditional RatSLAM algorithm.The comparison results show that the algorithm of this article has better advantages in constructing map accuracy,real-time performance and adaptability to the environment. |
