The Creation And Application Of Smart Car Cognitive Map

As a key technology of automatic driving system,SLAM has important research significance.The visual autonomous navigation system that imitates human cognitive behavior has become the research trend of SLAM.Aiming at the problem of location and navigation in urban scenes,this paper proposes a visual map-based algorithm to establish a cognitive map with decimeter-level positioning accuracy.It combines accurate location information with skilled driver's experience information to form an anthropomorphic Driving decisions and automating driving of complicated intersections.This paper proposes a DH-SSD target detection method and a modified BowRatSLAM map location system to solve the problem of occlusion of urban roads due to the obstruction of dynamic targets.The main contributions of this article are as follows.Aiming at the demand of real-time location maps for smart vehicles,this paper proposes a map location method based on bio-cognitive nervous system.This method inputs RGB image information and inertial mileage information into a virtual location memory network.Aiming at characteristics of many complex landmarks of urban environment,this paper proposes the BowRatSLAM model in the framework of traditional cognitive maps.The model uses ORB feature extraction to replace the original pixel-level matching,using the bag-of-words model to accelerate the image search,through training and testing in the KITTI automatic driving library and,a large number of experiments show that BowRatSLAM compared to traditional cognitive maps,positioning accuracy has improved significantly.For the problem of scene occlusion caused by dynamic targets in urban environments,this paper uses target detection to optimize the bag-of-words matching model to improve the positioning accuracy of cognitive maps.This paper first proposes a deep learning target detection method called DH-SSD.This method combines extended convolution and deconvolution to improve the SSD detection network,and has greatly improved the detection accuracy of outdoor small targets.This article uses the DH-SSD detection algorithm to detect dynamic targets(such as cars and pedestrians)and segment the dynamic obstacles in the image.Segmented image features can reduce error by visual closed-loop correction.The bag-of-word model performs vocabulary statistics on the segmented image feature points and accelerates the image search using the keyframe map library to achieve the purpose of improving the accuracy of the closed-loop correction.The test in the KITTI database shows that The 100-meter positioning error of this algorithm is improved by 1.95% compared to the RatSLAM algorithm.In this paper,the self-driving application is implemented in smart cars.Firstly,the above algorithm framework is used to create the cognitive maps.This paper adopts Dijkstra algorithm to realize the shortest path planning of cognitive map from any point to point.For smart cars in the face of complex urban environments,it is difficult to accurately locate and identify complex road conditions.This paper proposes to use cognitive map location information to provide skilled driver experience information for decision modules,and to realize anthropomorphic driving based on empirical control information.It solves the problem that it is difficult to achieve automatic driving at intersections,and realizes automatic driving test on smart vehicles.It achieves a 90% intersection driving success rate when testing real cars in the urban scene of Changchun.