Research On Image Local Feature Description Method For Visual SLAM

Visual SLAM,as the basis for smart devices,mobile robots,augmented reality and autonomous driving,is a hot topic in the field of computer vision.The purpose of visual SLAM is to estimate the location and orientation of the camera while reconstructing 3D maps of the environment.The inter-frame data association based on visual characteristics is the key point of visual SLAM system.Therefore,the extraction of the visual characteristics is one of the key issues of visual SLAM as visual features' robotness matters a lot.In this paper,we focus on the local feature descriptors for specific visual SLAM system and design a deep neural network for local feature descriptors.We propose an efficient SLAM system that uses deep local feature descriptors obtained by the neural network as substitute for traditional hand-caft features.The main work and contributions of this paper are as follows:1.The design of a high-performance descriptor network for specific applications,i.e.visual SLAM.There are three characteristics in the proposed neural network.First,the shallow convolution network structure is adopted to reduce the amount of calculation and ensure the running speed of the extraction process.Second,we improve the sampling strategy of training data to enable more training data to optimize the network.Third,the loss function is optimized by adding global orthogonal regularization term based on the traditional triplet margin loss to improve the prediction ability of neural network.Deep descriptors based on our proposd network are better than traditional ones and other deep descriptors based on the same network structure.In the meanwhile,the extraction consumes a short time,which can achieve real-time under GPU.2.For visual SLAM system based on deep feature descriptors,a uniform feature sampling strategy that combines adaptive threshold FAST feature point detection and quadtree filtering is proposed to meet the needs of a particular visual SLAM system.In the meanwhile,a training dataset of neural network is constructed according to the uniform feature sampling strategy.The proposed sampling strategy can ensure that the feature points are evenly distributed in the whole image.And the uniformly distributed feature points have an important role in promoting the visual SLAM system.Moreover,since the deep feature descriptors replaces the traditional ones in the visual SLAM system,it is necessary to use the training dataset based on the same uniform feature sampling strategy to maintain the consistency of data distribution between the training data and forecast data.However,the existing datasets' construction methods do not meet this requirement.Therefore,this paper uses the uniform feature sampling strategy to generate about 105 patch of 64 ×64 pixel size based on HPatches image sequences.The experimental results show that the network based on the uniform feature sampling strategy used in this paper outperforms the network trained on the traditional patch dataset in the visual SLAM system.3.The visual SLAM system is improved based on the deep descriptor designed in this paper.First,the traditional descriptor is replaced with deep descriptor based on the proposed network in the visual SLAM system.And the inter-frame data association is constructed based on deep descriptors.Second,we adjust the matching method of the feature matching module according to the characteristics of deep descriptor in this paper.Finally,the visual dictionary model based on the deep feature is used for modules such as relocation and loop closing of the visual SLAM system.The classical framework ORB-SLAM2 is selected to improve and test.The enhanced system achieves a great improvement and shows the most significant improvement on the Difficult sequence of EuRoC dataset.The systembased on deep descriptors is increased by an average of 31.6%in RMSE.