Research On Semantic Loop Closure Detection Algorithm In Complex Environment

As a basic component module in the field of Simultaneous localization and mapping(SLAM),loop closure detection is a key technology for the fields of unmanned driving,navigation planning and artificial intelligence,and has broad development prospects.Although many innovative achievements of visual loop closure detection algorithms were acquired in in recent years,current visual loop detection algorithms still face two main problems in complex environments: the interference of dynamic objects which affects the accuracy of detection and the change of light which reduces the recall rate of the loop closure detection.Faced with loop closure detection tasks in complex environments,the design process of semantic loop closure detection algorithm in this paper is as follow: Firstly,the panoptic segmentation technology which extracts semantic information is researched in our work.The reliability of semantic information of landmark is improved by designing the novel loss function which based on attention mechanism.Then,the method of semantic fusion in loop closure detection algorithm is explored,and a semantic loop detection algorithm based on multi-information fusion is proposed to improve the robustness of loop closure detection.The main work of this paper is as follows:(1)Aiming at the problem that the importance of semantic information of objects in different application scenarios is various,an attention-based semantic enhancement loss function is designed in this paper.This loss function improves the segmentation network’s attention to landmark by setting semantic hierarchical division of different object classifications and importance perception coefficients.At the same time,a panoptic segmentation network based on the feature pyramid network structure is built to improve the accuracy of the segmentation of important landmarks.(2)To reduce the influence of interference of dynamic objects and light changes in complex environments,a semantic loop detection algorithm based on multiinformation fusion is proposed in this paper.In this algorithm,a new type of semantic bag-of-words model which combines semantic labels with visual features is designed.The interference caused by dynamic objects is removed by this model.At the same time,a semantic landmark vector model is established to encode the semantic topology map.The geometric structure is encoded to improve the robustness of the system in a changing light environment.Different from traditional visual algorithms and detection algorithms based on deep learning,this algorithm combines semantic,visual and geometric information to improve the robustness of visual loop detection in complex environments at the same time.(3)A semantic SLAM system experimental platform is built,a matching method between the positioning module and the semantic loop detection algorithm is designed.It improves the accuracy of SLAM system.The system is tested in public data sets and practical environments.The experimental results show that the semantic loop detection algorithm based on multi-information fusion proposed in this paper can effectively cope with complex environmental changes and improve the robustness of the visual loop closure detection algorithm and SLAM system.