基于目标检测网络的动态场景下视觉SLAM优化  

Vision SLAM Optimization in Dynamic Scene Based on Object Detection Network

作　　者：方娟 方振虎 FANG Juan;FANG Zhenhu(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124

出　　处：《北京工业大学学报》2022年第5期466-475,共10页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61202076);北京市自然科学基金资助项目(4192007)。

摘　　要：为了降低动态环境对同时定位与建图(simultaneous localization and mapping,SLAM)位姿估计的干扰,提出一种将目标检测网络与ORB-SLAM2系统结合的方法.在帧间估计阶段,使用目标检测网络获取当前帧的语义信息,得到潜在可移动物体边界框,结合深度图像并根据最大类间方差算法分割出边界框内前景,把落在前景中的动态特征点剔除,利用剩下的特征点估计位姿.在回环检测阶段,利用边界框构建图像语义特征,并与历史帧比较,查询相似关键帧,与视觉词袋法相比,该方法查询速度快,内存占用少.在TUM Techni数据集上进行测试,结果表明该方法可以有效提高ORB-SLAM2在高动态场景中的性能.To reduce the interference of dynamic environment on the pose estimation of vision simultaneous localization and mapping(SLAM),a method to combine object detection network with ORB-SLAM2 system was proposed.In the inter frame motion estimation stage,the bounding box of potential movable objects was obtained by using object detection network to acquire the semantic information of the current frame.Combined with the depth image and according to the maximum between-class variance algorithm,the foreground in the bounding box was segmented,the dynamic feature points in the foreground were deleted,and the remaining feature points were used to estimate the pose.In the loop closure detection stage,the bounding box was used to construct image semantic features,and query similar key frames compared with historical frames.Compared with Bag of Visual Word,the method has faster query speed and less memory consumption.The method on TUM dataset was evaluated,and the results show that the proposed method can effectively improve the performance of ORB-SLAM2 in high dynamic scene.

关 键 词：同时定位与建图(SLAM) 动态场景 目标检测 图像分割 回环检测 位姿估计 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]