Visual simultaneous localization and mapping algorithm for dim dynamic scenes

Visual simultaneous localization and mapping (VSLAM) is invaluable in applications like autonomous driving, underground robot exploration, and robot search and rescue operations. However, traditional VSLAM struggles in dim dynamic environments, prompting researchers to develop specialized algorithms for such environments. Despite these efforts, many existing VSLAM algorithms often fail to meet real-time requirements.

In this study, a real-time VSLAM algorithm for dim dynamic scenes is proposed. First, this study uses an algorithm based on the Retinex theory to enhance the brightness of input red green blue(RGB) images. Because images captured by the red green blue-depth(RGB-D) cameras cannot directly obtain illumination maps, this study transforms the illumination solving problem into an optimization problem. By developing an illumination cost function and taking its derivative, the derivative is yielded as zero, and a photometric graph is obtained. Second, this paper adds an object detection thread on top of ORB-SLAM3 using YOLO to identify targets labeled “person.” Next, feature points in the YOLO detection frame are prefiltered using the Lucas-Kanade(LK) optical flow method. Further, the position of the feature point in the next frame is predicted using this method. If the distance between the actual and predicted positions is greater than the preset threshold, the feature point is set as a dynamic feature point, and the other feature points are set as static feature points. Additionally, Gaussian distribution analysis is used for secondary judgment. By analyzing the standard deviation of the depth of feature points in the YOLO box, the variance and standard deviation ratio of their depth that are greater than and less than the standard deviation are calculated, and the distribution of feature points is estimated. These metrics are compared against thresholds, and different thresholds are selected for subsequent standardized scores. Finally, the standardized score for each preset static feature point is calculated to evaluate the degree of deviation from the original data point, yielding a mean of 0 and a standard deviation of 1. In static scenes, the depth values of most feature points should be relatively consistent, and even if there are some deep changes, these changes should be distributed around a mean. The depth value of dynamic feature points may significantly deviate from the mean value of static feature points in the boundary box, resulting in higher standardization scores and larger square values than the set thresholds. Through this judgment method, all dynamic feature points are successfully selected and culled.

Test results on the TUM and Bonn public datasets show that the proposed algorithm reduces the absolute trajectory error by at least 86.93% in high dynamic environments and 27.61% in low dynamic environments compared with ORB-SLAM3. Among similar VSLAM algorithm types, the proposed algorithm offers faster processing speeds while maintaining high precision and achieves a good balance between accuracy and real-time performance. The actual scene verification shows the excellent performance of the proposed algorithm in dim scenes. In summary, through advanced image processing and accurate dynamic point culling, the proposed algorithm addresses the shortcomings of traditional VSLAM algorithms and achieves accurate positioning in dim dynamic environments.