基于注意力模型的视觉/惯性组合里程计算法研究

由于外界环境的干扰和传感器精度的限制,视觉/惯性组合里程计的输入数据存在一定的噪声,这会增加里程计的解算误差,而且误差会随着时间积累。针对以上问题,设计了一种基于注意力模型的视觉/惯性组合里程计算法。该算法使用卷积神经网络和长短时记忆网络分别构建了视觉特征提取器与惯导信息特征提取器,同时引入了两种注意力模型:加权组合网络以及开关组合网络,对视觉特征信息和惯导特征信息的融合噪声进行降噪处理。通过在组合里程计算法中添加闭环校正环节,有效地抑制了里程计误差随时间的积累。对比实验结果表明,设计的组合里程计算法与其他算法相比,无论在性能上还是在精度上都有明显的提升。

Research on Visual/Inertial Integrated Odometry Algorithm Based on Attention Models

Due to the disturbance of the external environment and the limitation of the sensor accuracy, there is noise in the input data of the visual/inertial integrated odometry, which will increase odometry error, and the error will accumulate with time. To solve the above problems, this paper proposes an attention model-based visual/inertial integrated odometry algorithm. This algorithm uses convolutional neural network and long short-term memory network to build visual feature extractor and IMU information feature extractor. At the same time, two attention models, soft attention and hard attention network, are introduced to reduce the fusion noise of visual and inertial feature information. By adding loop closure optimization in the integrated odometry algorithm, the accumulation of odometry error with time is effectively restrained. The experimental results show that compared with other algorithms, the proposed algorithm has a significant improvement in performance and accuracy.