Optimization Techniques For Depth Performance Of Distance Sensors In Mobile Robot Navigation In Indoor Scenarios

Distance estimation is an important capability for mobile robots and driverless cars,which need to be able to drive safely without colliding with anyone or anything,i.e.to achieve autonomous obstacle avoidance and navigation.The current mainstream distance sensors are Li DAR and cameras,but both are influenced by their respective operating principles and other factors,and both have shortcomings in depth acquisition.To address the problem of depth acquisition of distance sensors in mobile robots and driverless cars,this paper proposes a laser point cloud up-sampling method and an MRF depth map restoration method based on semantic segmentation,respectively.In order to improve the accuracy of semantic segmentation in indoor scenes,this paper proposes a deep deformable attention module for RGB-D data,which fuses the features of RGB images and depth images.To address the problem of low resolution of the vertical direction of the acquired point clouds and the missing point clouds due to the occlusion between objects,this paper proposes an up-sampling method based on the spatial characteristics of the point clouds and a linear fitting to complete the missing point clouds.The missing point clouds are filled by linearly fitting the point cloud to each laser beam in the missing region,interpolating the fitted curve to complete the missing point cloud,and interpolating the completed point cloud between adjacent laser beams in an angle-averaged manner to up-sample the point cloud.The method is able to set an arbitrary up-sampling factor,effectively increasing the density of the point cloud,with at least 62.3% and 124.6% of the number of points being up-sampled in one and two up-samples respectively.The missing point completions fit well with the original point cloud,and the number of points from a single up-sampling of the completed point cloud is increased by 4.7% compared to the number of points from a single upsampling of the uncompleted point cloud.The newly generated points meet the fluctuating characteristics of the LiDAR point cloud and the results are more reliable.For indoor scenes due to uneven illumination,many types of objects,small spatial distance between objects and similar color and texture between objects,resulting in a huge challenge in accurate semantic segmentation of indoor environments relying solely on color images.We propose a depth-deformable attention module that fuses depth information with visual information in the backbone network feature extraction stage to obtain bimodal features.The sampling offset nature of deformable convolution is used to improve the influence of traditional convolution by fixed geometric structures,enhancing the adaptability of the convolution kernel to objects of different shapes and sizes,and the inclusion of an attention mechanism enables more effective feature extraction.Compared with the strategy of fusing high-level features and low-level features twice in the encoding stage,this method was tested on the NYUv2 dataset,with mIoU improving by 9.59% and 7.97%,PA improving by 4.87% and 3.35%,and MPA improving by 15.64% and 13.38%,respectively,and also has significant advantages compared with other advanced methods.To address the problems of depth voids and blurred edge depths in the depth map acquired by the depth camera,this paper proposes a depth missing processing and depth map smoothing and noise reduction method based on semantic segmentation.Firstly,the color image and the depth image aligned with the color image are acquired by the sensor,and the color image and the depth image are fed into the semantic segmentation network to segment the depth-smoothed individual objects;then,the spatial and structural correlations between the depth data and the color data are used to fill the depth voids in combination with the semantic segmentation result labels;finally,a Markov random field(MRF)model is constructed and combined with the semantic segmentation result labels,the depth map is repaired.The experimental results show that the algorithm in this paper has better robustness to different depth voids compared with traditional MRF algorithm and bilateral filtering algorithms,and the PSNR value of the restored image is improved by at least 7.7% in the case of complex depth voids.The depth map of real scenes is restored by the algorithm in this paper,the image is overall smooth,the edges are clear,artifacts are improved and the image quality is significantly improved.This paper focuses on the depth representation optimization of two depth acquisition sensors,Li DAR and depth camera,as well as indoor RGB-D semantic segmentation.The proposed method is important for 3D reconstruction,reverse engineering,indoor mobile robotics,autonomous driving perception and other fields.