【摘要】 为应对传统视觉惯性SLAM系统在动态环境中因静态世界假设失效而导致的定位精度下降和鲁棒性差，提出了一种前后端协同的动态鲁棒视觉惯性SLAM方法。在前端，引入了基于运动一致性的概率性识别框架，通过对比观测视差角与惯性测量单元（Inertial Measurement Unit,IMU）预积分值，结合一阶时序滤波，为每个特征点生成平滑、连续的动态置信度，对高置信度动态特征予以初步剔除或软抑制；在后端，设计了一种融合该动态置信度先验的自适应鲁棒优化器，将前端概率先验信息引入自适应截断最小二乘（Adaptive Truncated Least Squares,ATLS）代价函数框架下的加权光束法平差（Bundle Adjustment,BA），从而实现对动态外点的分层抑制与滤除。在VIODE(Visual-Inertial Odometry in Dynamic Environments)数据集上进行的实验表明：与DynaVINS算法相比，所提方法在高动态序列city＿day＿high和parking＿lot＿high上的绝对轨迹误差分别降低了29.0%和38.6%，而且随着动态程度的增强，性能下降更为平缓，展现出更强的鲁棒性。实验结果表明，该方法能够有效提升在复杂动态环境下视觉惯性SLAM系统的精度与可靠性。更多还原

【Abstract】 Traditional visual–inertial SLAM systems suffer from degraded positioning and poor robustness in dynamic scenes, as the staticworld assumption is no longer valid. To address this problem, a collaborative front-end/back-end VI-SLAM approach is proposed that explicitly handles moving features. At the front end,a probabilistic identification framework based on motion consistency is introduced. By comparing the observed parallax angle with the pre-integrated value from the inertial measurement unit(IMU) and combining first-order temporal filtering, smooth and continuous dynamic confidence values are generated for each feature point. Features with high confidence are treated as likely moving and are either removed early or softly down-weighted. In the back end, an adaptive robust optimizer fused with the prior dynamic confidence is designed. The probabilities prior information from the front end is introduced into the weighted bundle adjustment(BA) under the framework of the adaptive truncated least-squares(ATLS) cost fuction, so that staged weighting suppresses and dynamic outliers are realized. Experiments conducted on the VIODE(Visual-Inertial Odometry in Dynamic Environments) dataset show that,compared to DynaVINS, the proposed method reduces the ATE(Absolute Trajectory Error) RMSE(Root Mean Square Error) by 29.0% and 38.6% on the high-dynamic sequences city＿day＿high and parking＿lot＿high, respectively. Furthermore,as the degree of dynamics increases, the performance of this method degrades more gradually, demonstrating superior robustness. The experimental results confirm that the proposed method can effectively enhance the accuracy and reliability of VI-SLAM systems in complex dynamic environments.更多还原