复杂环境下基于自适应极线约束的AGV视觉SLAM算法

Adaptive polar constraint-based visual SLAM algorithm for AGV in complex environment

针对传统视觉同步定位与地图构建(SLAM)算法不能有效处理复杂环境中的动态及潜在动态目标而影响定位与建图性能的问题,提出一种基于Mask R-CNN神经网络以及ORB-SLAM3算法改进的视觉SLAM方法。针对动态目标,提出一种基于语义信息的运动一致性检验算法,使用自适应阈值的极线约束方法实现图像中动态特征点的精确剔除;针对潜在动态目标,提出一种改进的长期数据关联方法,通过增大关键帧选取密度及优化关键帧中的潜在动态目标信息,对算法的回环优化和地图融合过程进行改进,提高回环优化效果与地图复用性。在TUM数据集和真实场景中进行验证,实验结果表明与ORB-SLAM3算法相比,采用TUM数据集在低动态场景和高动态场景中的绝对轨迹均方根误差分别减小8.5%和65.6%;在真实场景下测试,所提算法的定位精度提高了62.5%。

Aiming at the problem that traditional visual simultaneous localization and mapping(SLAM)algorithm cannot deal with dynamic and potential dynamic targets in complex environment effectively,which affects the localization and mapping performance,an improved visual SLAM method based on Mask R-CNN neural network and ORB-SLAM3 algorithm is proposed.A semantic information-based motion consistency checking algorithm is proposed for dynamic targets,using an adaptive thresholding polar line constraint method to achieve accurate rejection of dynamic feature points in images.An improved long-term data association method is proposed for potential dynamic targets.By increasing the key frame selection density and optimizing the potential dynamic target information in key frames,the loopback optimization and map fusion process of the algorithm are improved to enhance the loopback optimization effect and map reusability.The experimental results show that compared with the ORB-SLAM3 algorithm,the proposed algorithm reduces the root mean square error of absolute trajectory in the low dynamic scenes and high dynamic scenes of TUM data set by 8.5%and 65.6%respectively.The localization accuracy of the proposed algorithm has 62.5%improvement compared with ORB-SLAM3 in real scenario.