Semantic Topological Map Based On Spatial Cognitive Model Of Brain

Simultaneous Localization and Mapping(SLAM)has always been a hot issue in the field of robotics.In the interaction and perception with the environment,the service robot needs memory function to complete the cognitive task,and the RatSLAM based on the brain spatial cognitive model has this characteristic.RatSLAM can describe and remember its location just like a human does,using the iconic features of the scene,without the need for high-precision sensors,and is suitable for long and wide range positioning and navigation.However,the positioning and mapping accuracy of RatSLAM is not high in complex environments,such as significant lighting changes,fuzzy motion,and untexturing areas.At the same time,the cognitive map of RatSLAM is a two-dimensional topological map,so the robot cannot understand the environment from a higher semantic level.In this paper,a semantic topology map construction method based on spatial cognitive model of brain is proposed based on RatSLAM algorithm.Firstly,for the original RatSLAM algorithm uses the sum of absolute difference(SAD)template matching algorithm accuracy is not high,the shortcomings of sensitivity to noise,using differential hash algorithm(dhash)get the current image and the hash value of the view template and calculate the corresponding hamming distance,if the hamming distance is less than the setting threshold,then on to the next step based on the movement of the grid statistics feature matching(GMS).GMS feature matching algorithm has both robustness and real-time performance,which can further improve the accuracy of template matching,so as to modify the active unit of pose cells more accurately and obtain more accurate experience map.Laboratory experiments show that the proposed algorithm ensures real-time performance,and the matching accuracy of view template and empirical map is higher than that of the original RATSLAM method,which can form accurate closed-loop.Secondly,the robotic-centric visual inertial odometry(R-VIO)was introduced to solve the problem of poor positioning accuracy of pure visual odometry based on the original RatSLAM algorithm in complex environments,such as violent movement,weak texture and obvious illumination changes.It integrates the information of camera and inertial measurement unit(IMU)sensors.With the traditional algorithm for the center with the mobile robot is estimated directly relative to the global reference frame fixed absolute motion,R-VIO estimate relative to the movement of the more accurate relative movement of local reference frame,to improve the positioning accuracy of robot in complex environment,to ensure that the posture cells of robot pose estimation accuracy,and,in turn,generate more accurate map of experience.In this paper,the reliability of the proposed algorithm is verified on all sequences of Eu Ro C dataset.Experiments show that the trajectory error of the proposed algorithm(RVIO-RatSLAM)is less than that of the original RatSLAM method,and it can form a more accurate empirical map.Finally,the original RatSLAM empirical map is just a simple topological map,which cannot understand the surrounding environment from the semantic level.Therefore,the Yolo-v3 target detection network is introduced to obtain the semantic labels of the objects in real time and accurately.Because the problem that Yolo-v3 cannot extract semantic labels in the case of visual occlusion,Yolo-v4 network is further adopted to extract semantic labels of occluded objects and map semantic labels into the topological nodes to generate the semantic topological map.This algorithm also has the function of memory.According to the correlation between view template and pose cell,it can judge whether the detected object position has changed,and the topology node will be updated with the change of the object.Experimental results on Oxford?newcolleage dataset and real environment show that the algorithm can construct clear and accurate semantic topological map,and understand the environment more accurately and deeply.