| Simultaneous Localization and Mapping(SLAM)is regarded as the key technology to realize autonomous localization and navigation of mobile robots.It is also a research hotspot in the fields of unmanned driving,augmented reality and virtual reality.With the enlargement of the mobile robot application scope and the complexity of application scenarios,the accuracy and robustness of traditional SLAM algorithms have been severely challenged.For example,sparse features can cause initialization or tracking failures in low texture or single texture scene.The sensors will have the accumulating error,which can not realize accurate loop closure detection in large scale environments.The algorithm is easy to track failure and converge to the error value in complex environments such as illumination,seasonal and perspective change.In order to solve the above issues,this paper deeply studies the relevant algorithms of SLAM system and proposes improved algorithms from three aspects of motion estimation for SLAM front end,optimization for SLAM back end and loop closure detection.Experimental results show that the proposed algorithm significantly improves the robustness and efficiency,and can satisfy the requirements of practical applications.The main contents are as follows:To solve the problem of initialization or tracking failure in the sparse texture environment,the motion estimation algorithm based on feature fusion is proposed.The improved random sampling maximum likelihood algorithm is used to extract planar features,which eliminates the dependence of traditional algorithm on empirical thresholds.The parallel,perpendicular and planar point constraints of planar structures are introduced to reduce the accumulating drift.The camera motion is estimated by minimizing the constraint function based on the geometric constraint relationship of features.Experimental results show that the proposed algorithm has high accuracy and wide application compared with ORB SLAM and LSD SLAM.To reduce the accumulating drift of sensors in large scale environments,the back end optimization algorithm based on fusion of improved heuristic algorithm and particle filter is proposed.The nonlinear inertial weight is introduced to improve the optimization precision and speed up the convergence in the position and speed updating of heuristic algorithm.The penalty function and exclusion mechanism of niche are used to effectively ensure the diversity of population.The improved heuristic algorithm is used to optimize the sampling process of particles,which makes particles move to the high likelihood region and improves the global optimization ability.Experimental results show that the proposed algorithm effectively solves the problems of particle degradation and particle poverty,significantly improves the positioning precision and the real time performance of mobile robots.To improve the accuracy of location recognition in complex environments,a hybrid method of quaternion and convolutional neural network for loop closure detection is proposed.The multi-scale landmarks are extracted by superpixel segmentation to make image description have the appearance invariance and viewpoint invariance.The convolutional layer is extended to the quaternion spatial convolutional layer,which effectively handles the coupling between red,green and blue channels in color images,extracts the deep information of color images,and reflects the integrity of color images.Combined with the spatial pyramid pooling model to suppress overfitting and increase model invariance.Not only the distance of matching landmarks but also the spatial distribution and shape constraint information of landmarks are considered in similarity measurement.The experimental results show that the proposed algorithm not only guarantees a high recall rate,but also improves the accuracy of loop closure detection. |
PDF Full Text Request