Research On Multi-feature Fusion And Closed-loop Detection Location Method Of Visual SLAM For Intelligent Vehicle

Visual SLAM(Simultaneous Localization and Mapping)technology plays an important role in the autonomous positioning technology of intelligent vehicles.However,vehicle localization performance is still poor even with state-of-the-art visual SLAM methods in some road environments,such as icy and snowy roads,rainy and foggy weather.To improve the comprehensive localization performance of SLAM systems,a closed-loop visual SLAM method based on fusion of point and edge features is proposed.The main research contents are as follows:Visual SLAM cannot produce reliable camera motion estimation in an environment with weak texture areas,which is prone to tracking loss or failed localization.A visual SLAM method based on adaptive fusion of point and edge features based on the improvement of the ORB-SLAM2 system framework is proposed.In the front-end of visual SLAM,it is judged whether the current vehicle’s driving environment is a weak texture environment or not by evaluating the quality of the feature points in the image.In the case of weak texture environment,edge features are introduced to construct joint error constraints of point and edge to optimize the camera pose.At the same time,the rotation constraints are added to improve the keyframe selection strategy.Also,the distance-transform method is used in the pose optimization of edge features,which can effectively improve the speed of iterative optimization.The accuracy of closed-loop detection based on the bag-of-words become low in an environment with high similarity.A loop closure detection method based on bag-of-words and global descriptors is proposed,which mainly by introducing the global descriptor of the image to enhance the constraint of image information and improve the accuracy of closedloop detection.Firstly,screen candidate frames by the gray histogram similarity between images.Secondly,the gray histogram and bag-of-words vector similarity scores between images are combined to obtain a weighted similarity score,and the candidate frames that fail the similarity threshold are filtered.Finally,the selected candidate frames are verified to meet the requirements through geometric verification which includes epipolar geometric constraints,time and space consistency test.The proposed method is evaluated based on the KITTI and TUM datasets.Compared with related visual SLAM algorithms,the method proposed in this paper has better positioning accuracy and robustness in weak texture scenes.And it has only a little increase in time consumption,which ensures the real-time performance of the system.The closedloop detection method proposed in this paper is evaluated by evaluation indicators such as precision-recall and positioning accuracy.The experimental results show that the method proposed has a significant improvement in system positioning accuracy,and has a higher recall rate under 100% accuracy.Through joint testing,the feasibility of combining the two methods proposed to simultaneously improve the performance of visual SLAM localization is explored.