河北软件职业技术学院学报

为提高视觉SLAM算法在弱纹理动态场景的定位准确性，设计了一种基于点线特征融合和增加目标检测线程进行图像检测的视觉SLAM优化方法。该算法基于ORBSLAM2开源框架，在特征提取阶段增设了线特征提取功能，从而能够同时提取点线特征，获取更为丰富的特征信息；增加目标检测线程，利用MNN推理引擎搭载MobileNet-SSD模型进行动态对象检测，使得算法能够检测出输入图像的动态特征，结合极线约束的几何方法将动态特征点剔除，结合点线特征与动态检测结果，实现了SLAM算法在弱纹理动态场景中快速且精准定位。通过一系列的对比实验，验证了算法在对应场景的性能，提高了SLAM算法的定位准确性。

To improve the positioning accuracy of visual SLAM algorithms in weak-texture dynamic scenes, an optimized visual SLAM method integrating point-line feature fusion and an additional target detection thread for image detection is proposed. Designed based on the opensource ORB-SLAM2 framework, the method adds a line feature extraction function in the feature extraction phase,enabling simultaneous extraction of point and line features to obtain richer feature information. It also introduces a target detection thread, which uses the MNN inference engine with the MobileNet-SSD model for dynamic object detection,allowing the algorithm to detect dynamic features in input images. By combining geometric methods with epipolar constraints to eliminate dynamic feature points, and integrating point-line features and dynamic detection, the SLAM algorithm achieves fast and accurate localization in weak-texture dynamic scenes. A series of comparative experiments verify the performance of the algorithm in relevant scenarios, improving the positioning accuracy of the SLAM algorithm.