Research On V-SLAM Algorithm Based On Multi-sensor Information Fusion

The ability of self localization and environment construction is an important prerequisite for mobile robots to complete tasks autonomously,and is the basis of upper application.Vision sensor has been widely used in the field of robot because of its ability to obtain rich environmental information and its low price.For Visual SLAM(Visual-Simultaneous Localization and Mapping,V-SLAM for short),feature point method is the main method at present.Although the positioning speed is fast,it can not work stably when dealing with the scene with fast rotation and missing features.Compared with point features,there are abundant line features in artificial environment,which can effectively alleviate the dependence of the system on point features in weak texture scenes,and avoid the situation of system state estimation error caused by single features or missing features.Compared with the pure vision SLAM method,the typical vision-inertial odometer scheme provides higher precision and more robust positioning services.However,there are still many areas to be improved in both pure vision and vision-inertial schemes.In view of the high accuracy of robot positioning and the difficulty of high availability of the built map,this paper makes the following improvements from the extraction of image features and the calibration of multi-sensor external parameters:1.On the basis of traditional on-line calibration of external parameters,in addition to the inherent geometric constraints,the point re projection residual and IMU measurement residual constraints are added to obtain a higher accuracy of external parameter estimation.In addition,as for the traditional on-line calibration algorithm of external parameters and off-line calibration,it is impossible to detect whether the external parameters change in time when the system is running.In this paper,the external parameter calibration ring is used The node is coupled with the positioning link,and the external parameter value is always regarded as one of the optimization variables,and the external parameter value is used as the optimization variable in the whole process,so that the system can reduce the impact of positioning accuracy decline caused by the change of external parameter during long-time operation..2.Aiming at the shortcomings of LSD line feature extraction algorithm that the quality of extracted features is uneven,a definition about the length of line features is proposed,so that the slam positioning system can only use long line features with more information and avoid using short line features that are even degenerated to "point features",so as to achieve higher positioning accuracy and faster data association than the original LSD algorithm In view of the vision inertial odometer integrating point and line features,in order to avoid feature redundancy,this chapter further removes line features according to the density of point features,so as to further reduce the number of feature extraction,so as to reduce the time of feature matching without affecting the positioning accuracy.3.To solve the problem of sparse mapping in feature-based slam system,this paper uses Scharr filter to get image gradient,and extracts pixel points higher than threshold value as the road marking points used by the mapping link alone,instead of participating in location estimation,so as to realize the feature-based slam which can not only locate quickly but also complete dense mapping.The validity of the improvement is verified on the real dataset.