A Computationally Efficient Semantic SLAM Solution for Dynamic Scenes

Abstract

In various dynamic scenes, there are moveable objects such as pedestrians, which may challenge simultaneous localization and mapping (SLAM) algorithms. Consequently, the localization accuracy may be degraded, and a moving object may negatively impact the constructed maps. Maps that contain semantic information of dynamic objects impart humans or robots with the ability to semantically understand the environment, and they are critical for various intelligent systems and location-based services. In this study, we developed a computationally efficient SLAM solution that is able to accomplish three tasks in real time: (1) complete localization without accuracy loss due to the existence of dynamic objects and generate a static map that does not contain moving objects, (2) extract semantic information of dynamic objects through a computionally efficient approach, and (3) eventually generate semantic maps, which overlay semantic objects on static maps. The proposed semantic SLAM solution was evaluated through four different experiments on two data sets, respectively verifying the tracking accuracy, computational efficiency, and the quality of the generated static maps and semantic maps. The results show that the proposed SLAM solution is computationally efficient by reducing the time consumption for building maps by 2/3; moreover, the relative localization accuracy is improved, with a translational error of only 0.028 m, and is not degraded by dynamic objects. Finally, the proposed solution generates static maps of a dynamic scene without moving objects and semantic maps with high-precision semantic information of specific objects.

Feature-based visual simultaneous localization and mapping: a survey

“…A multitude of different SLAM approaches were proposed based on the use of a combination of low-and high-level features in [10,26,37,48,71,105,106,118,130]. Such approaches demonstrate high-level expressiveness while maintaining robustness.…”

“…The remaining data image a static environment which can be processed using a standard visual SLAM algorithm. A similar approach can be found in [130] where dynamic objects are segmented out of the scene by means of a computationally efficient step-wise approach to detect the object and extract its contour. The static environment is then mapped based on point features using a novel look-up table approach that targets using a large amount of distinct, evenly-distributed point features from the environment, which enhances the accuracy of mapping and localization.…”

SDF-SLAM: Semantic Depth Filter SLAM for Dynamic Environments

“…By applying the classic semantic segmentation networks, such as YOLO [8], SSD [9], Seg-Net [10], Mask-RCNN [11], PSPNet [12], and Deeplab [13], the semantic labels of the extracted image features in visual SLAM framework can be obtained. When the objects in the image are recognized as movable objects, such as people, cat, and car, the features located on these objects are thought as dynamic features and will be directly removed [14]- [16] or further processed through a selective tracking method in the tracking thread of SLAM [17] to determine whether they are retained or removed. The idea of using semantic information to detect dynamic feature points is very simple and direct, but it also has some limitations, mainly including two aspects: first, the semantic dynamic feature points do not completely coincide with the actual dynamic ones; second, the semantic segmentation results have errors especially in the boundary region of objects.…”

SOF-SLAM: A Semantic Visual SLAM for Dynamic Environments

“…Zhang et al [22] use YOLO to get semantic message, they consider features which are always located on the moving objects as unstable and filter them out. Wang et al [23] propose a step-wise approach that consists of object detection and contour extraction to extract semantic information of dynamic objects in a more computationally efficient way. Xiao et al [24] use SSD object detection network running in a separate thread to get prior knowledge about dynamic objects, and the features on dynamic objects are then processed through a selective tracking algorithm in the tracking thread, to significantly reduce the error of pose estimation.…”