Research On Autonomous Navigation Of Intelligent Vehicle Based On Semantic SLAM

Simultaneous Localization and Mapping(SLAM)has been widely used in environment perception and autonomous navigation of robots and smart cars.Thanks to advances in camera technology and computing performance,visual SLAM has made great progress.At present,there are still the following problems to be solved in the visual SLAM system: On the one hand,the static assumptions that the visual SLAM research relies on are often difficult to satisfy,and dynamic objects not only affect the accuracy of the camera pose solution,but also lead to environmental map construction and positioning problems.Obvious deviation,which affects the performance of positioning and obstacle avoidance in subsequent autonomous navigation;on the other hand,traditional visual SLAM does not contain scene semantic information,which reduces the accuracy and stability of environmental perception.In view of the above problems,this paper proposes a complete semantic SLAM intelligent vehicle autonomous navigation system based on a lightweight and efficient semantic detection algorithm and an adaptive dynamic feature point discrimination method.The main research contents are as follows:Aiming at the problem of lack of semantics in ORB-SLAM3,a lightweight and efficient semantic detection algorithm is proposed,which improves the ordinary convolution into a lightweight Ghost Conv,which eliminates a large amount of redundant information in the traditional deep learning network;adds replacement attention mechanism improves the attention to small targets;Bi FPN structure is introduced to strengthen the feature fusion of the backbone network and the neck network.Using semantic detection algorithm as a priori condition,a visual odometry based on adaptive dynamic feature point discriminant method is designed,which improves the performance of ORB-SLAM3 algorithm under dynamic conditions.When the adaptive dynamic feature point discriminant method receives the image frame containing the prior dynamic object sent by the semantic detection algorithm,the feature points are classified according to the number and percentage of the prior dynamic object,and the LucasKanade optical flow and RANSAC algorithm are used to perform Eliminate optimization.In view of the lack of semantic information of actual objects in the gridded grid map,which affects the accuracy and stability of the intelligent vehicle’s environmental perception,this paper proposes a grid map semantic enhancement algorithm based on ORB-SLAM3,and locates,recognizes and tracks the grid Objects of different categories in the map to generate a semantic raster map.The specific operation is: use the key frames and feature points obtained by visual odometry;use the Bresenham algorithm to draw the ray to connect the feature points,and then use the access counter and the occupation counter to record the state of each feature point in the grid;calculate the grid by the probability formula The grid occupancy probability is used to construct a grid map according to the threshold;the semantic detection and 3D point cloud matching results are projected onto the grid map to generate a semantic grid map,which provides the necessary map conditions for the subsequent autonomous navigation of smart vehicles.Aiming at the indoor scene of the smart car,the Turtlebot2 mobile robot platform was built,and the autonomous navigation system of the smart car was designed according to the semantic SLAM algorithm and the ROS operating system.Path planning and obstacle avoidance experiments were carried out in the simulation environment and the real environment,and the results verified that the smart car autonomous Availability of the navigation system.