Research Of Optimization Strategy For Monocular Vision Simultaneous Localization And Mapping

Mobile robot technology is increasingly becoming the focus of much attention.In the future,SLAM technology will not only integrate with autonomous vehicles to facilitate people's travel efficiency,but also combine with industrial transportation to make unmanned trucks and drones improve circulation efficiency.It will also be applied to VR and AR glasses to change people's interactive concepts of life and entertainment.These familiar scenes in our life will change dramatically in the future due to the advancement of SLAM technology.The current challenges of Visual SLAM are two-fold: first,how to maintain the stability of the algorithm system in a dynamic scene of a mobile robot's continuous moving perspective,second,how to effectively use the rich data information obtained from sensors and efficiently optimize calculation results under existing hardware equipment.Based on the above problems,this paper applies the Factor Graph model to the optimization algorithm of monocular vision,and proposes a monocular vision SLAM method of information fusion.In terms of current issues such as the visual odometry of SLAM in outdoor dynamic scenes has low utilization of feature point information in the visual front-end and insufficient to meet stable image matching and tracking.This paper proposes a front-end visual odometry information fusion method combining feature key and direct method gray-level matching method,which not only has no serious feature loss under fast motion conditions,but also reduces the dependence on gray-scale invariant hypothesis.It also has better stability in undertextured scenes.The proposed method also has good stability in the scenes with insufficient texture information.The experiment proves that the front-end visual odometry further simplifies the feature points,and then calculates the pixel gradient of the pixel block,which improves the utilization of the pixel information of the visual front-end image,improves the loss of feature points in the traditional method and reduces the dependence on gray-scale invariant hypothesis.At the same time,the average real-time frame rate meets the requirements of real-time performance.At the back end of the visual odometry,this paper abandons the traditional filter algorithm and constructs a factor graph model through recursive Bayesian estimation.After obtaining the sensor measurement information in the front end of the visual odometry,the factor graph is extended through the system's variable nodes and factor nodes,and the recursion and update of the state are completed based on the set cost function.Through the factor graph optimization,the incremental smoothing method is used to marginalize the pose matrix,so that the monocular vision SLAM can maintain the real-time requirements in outdoor large-scale scenes,and further improve the positioning accuracy.Experiments show that compared with the traditional SLAM algorithm,the average accuracy of the root mean squared error of the positioning accuracy of the visual odometry method proposed in the KITTI odometry datasets is 1.79 m.After adding the factor graph to optimize the model,the average root mean square error is 1.63 m.The rate reaches 71 fps.Experimental analysis shows that under the same hardware environment,compared with other traditional visual odometries,our algorithm has certain improvements in accuracy and stability.At the same time,based on the factor graph non-linear optimization model introduced in the visual odometry,the positioning accuracy of the visual odometry is further improved,and a higher frame rate level is ensured to meet the needs of real-time positioning.And has good stability.