| In recent years,due to the rapid development of information technology,artificial intelligence has ushered in a stage of rapid development.A series of artificial intelligence applications represented by intelligent robots have gradually penetrated into all aspects of people's lives.In order to complete the target oriented autonomous movement in the unknown environment,the intelligent robot must solve two basic problems of environmental perception and its own positioning.Simultaneous localization and mapping(SLAM)combines the two key issues of mobile robot positioning and map construction.General map construction and positioning systems require prior information of the environment or the model of the environment,but SLAM system can still work under the condition that the map information and the robot's position are not clear,and it can still work under the condition that the environment changes.SLAM method based on mathematical probability is based on the sensor platform.But creatures in nature take an entirely different approach to the problem,building maps and navigating complex environments without the need for high-precision sensing devices.This is because animals have the ability to synchronize map building and navigation.Inspired by the spatial cognition of animals,Australian scholar Michael et al.proposed a rodent hippocampal expansion model(RatSLAM).The principle and model of RatSLAM bionic navigation algorithm are introduced in detail.Since the formation and matching process of local scenes in the traditional RatSLAM algorithm adopts absolute difference value and model,there is no geometric processing and feature extraction of the scene,and its significant disadvantages are the sensitivity to the change of light and the inability to recognize the image after the rotation of the existing template image.The global feature can quickly complete the overall judgment of the scene,and the local feature can provide information such as the spatial distribution of objects in the scene,which is complementary to the global feature shape.Therefore,in the stage of local scene formation,GIST features and SIFT features of local scene images are extracted respectively.After serial fusion,local scene feature templates are formed and stored in local scene cells.The robot posture was modified by sensing the correlation between the local scene cells and posture cells.Real-time template matching was carried out through local scene cells to detect and correct the experience map.Through simulation experiments,the matching effects of RatSLAM model,RatSLAM+SIFT model,RatSLAM+GIST model and feature fusion RatSLAM model experience nodes,as well as the matching effects of RatSLAM model and feature fusion RatSLAM model experience map are qualitatively analyzed.The accuracy and recall rate of RatSLAM model and feature fusion model are analyzed quantitatively.Compared with the original RatSLAM model,the improved RatSLAM model improves the detection accuracy,improves the sensitivity of the system to light changes,and enhances the robustness of the system.The traditional RatSLAM algorithm performs closed-loop detection and relies entirely on visual input when correcting the experience map.When visual cells detect a familiar environment through scene matching,visual cell activity increases and visual activity is injected into the pose-aware cell matrix,causing the robot to reposition its predicted pose.Therefore,after extracting the GIST features and SIFT features of the local scene image,the visual word bag is trained separately,and the closed-loop detection is completed by the visual word bag,and the loopback verification link is introduced to improve the correct rate of the closed-loop detection.Through simulation experiments,the accuracy and recall rate of RatSLAM model based on visual word bag and visual word bag model with ring validation are analyzed quantitatively.Compared with the traditional RatSLAM algorithm,the accuracy and recall rate of the improved RatSLAM model are improved. |
PDF Full Text Request