Research On Motion And Environment Perception Method Based On Vision/inertial Combination

In this paper,a synchronous localization and mapping system based on deep learning neural network is proposed.The work is mainly divided into the design of visual-inertial odometry network based on deep learning and the research of monocular camera depth estimation algorithm.In the visual-inertial odometry network based on deep learning,the visual feature extractor based on convolution neural network and the IMU information feature extractor based on long short-term memory network are designed,and the window optimization network is designed to optimize the relative pose in short time.Aiming at the possible noise in the fused feature of visual feature and IMU information,this paper designs two kinds of attention networks,which are soft attention and hard attention.The attention network outputs a weight mask to filter the fused feature,and the dimension of weight mask is same as that of fused feature.In the research of monocular camera depth estimation algorithm,the depth estimation network and relative pose estimation network based on residual network are designed.In order to enhance the multi-scale features of depth image,the low-level features output by the feature extractor of depth estimation network are connected to the corresponding convolution layer of the depth decoder.Aiming at the outliers of the reprojection error,this paper introduces the concept of per-pixel minimum reprojection loss,and ignores the error value which does not conform to the projection model in the forward propagation stage,so as to further improve the performance of the network.Aiming at the problem that the error of learning-based visual-inertial odometry network accumulates with time,this paper adds the loop closure optimization to the odometry,and uses the graph-based optimization algorithm to correct the global pose.Finally,this paper uses volumetric integration algorithm,combining the pose and depth of image frame to build dense scene map.The experimental results show that the outdoor scene reconstructed by the synchronous localization and mapping system proposed in this paper has rich details and clear object contour.