Outdoor Mobile Robot Radar-image Cross-modal Retrieval Technology

With the continuous development of the economy and society,the level of science and technology is getting higher and higher,and the development level of mobile robots has been greatly improved.Many industries have begun to apply mobile robots to their daily work and life.In recent years,due to the rapid advancement of control science,computer and bionics,the working level of mobile robots has been improved,and the types and application scenarios of mobile robots have become more and more extensive.It has begun to gradually develop from simple indoor scenes to outdoor ones.Due to the complex and varied outdoor environment,for mobile robots working in outdoor environments,including the currently popular driverless cars,excellent environmental awareness and understanding performance can help them better navigate,locate,and successfully complete the intended goals.The radar-image cross-modal retrieval of outdoor mobile robots studied in this paper is part of the visual perception system of mobile robots and realizing radar-image cross-modal retrieval has important practical significance for the next environment awareness,self-positioning and navigation.At present,mobile robots mainly rely on point cloud data information collected by laser radar and image information collected by cameras to sense the surrounding environment.Due to the difference between point cloud data and images,In the case of extreme weather or night,the image captured by the camera will be greatly disturbed.So this paper first proposed a radar image cross-modal retrieval technology for outdoor mobile robots based on cluster canonical correlation analysis and deep learning network,firstly using deep learning network PointNetVLAD and AlexNet to extract the feature vectors of the point cloud and the image data respectively,and then use the canonical correlation analysis to map the feature vectors of the two modes to the same subspace.Finally,by calculating the Euclidean distance for retrieval,the image that matches the point cloud best can be obtained.The experiments show that cluster canonical correlation analysis is better than canonical correlation analysis in this paper.The retrieval method proposed in this paper is validated on the KITTI dataset,and the cross-modal retrieval from point cloud to image is realized.And the validity of clustering canonical correlation analysis in radar image retrieval of outdoor mobile robots is verified by the comparison of retrieval results and actual images.