| Image feature point detection is not only a basic task in computer vision,but also a key link in pose estimation tasks for unmanned platforms.However,with the escalation of current task requirements and application environments,especially in low-texture or no-texture scenes,the accuracy of traditional feature point detection algorithms is limited,which makes the performance of related point feature-based pose estimation algorithms poor.Although the feature point detection algorithm based on deep learning has made great progress,the balance between accuracy and efficiency still needs to be solved urgently.In this paper,based on the performance advantages of deep learning,the feature point detection algorithm for unmanned platform position estimation is studied with a favorable trade-off of accuracy-speed.The main research contents are as follows:1.Exploring feature similarity between network channels to propose a lightweight and efficient feature point detection algorithm.Firstly,an efficient stepwise convolution strategy is designed to decouple the feature coding methods between channels,which allows different feature encoding methods to be used for encoding to improve the efficiency of feature encoding.Then,according to the encoding characteristics of different stages of the network,three different types of multi-branch convolution modules are designed to enhance the feature extraction ability.Finally,a model compression method combining stepwise convolution strategy and structural reparameterization technology is proposed to improve the inference speed of the network by parameter fusion of multi-branch modules.Experimental results on the Hpatches dataset and real-world environments show that the proposed algorithm achieves comparable detection performance to the recent work with only half the network parameters of Super Point algorithm,demonstrating a favorable trade-off between accuracy and speed.Especially,compared with the DAN-Super Point and R2D2 algorithm,the feature point detection speed of 480×640 image is increased by 48.5% and 41.4%,respectively.2.Aiming at the practical application requirements of pose estimation tasks for unmanned platforms,a framework for pose estimation and sparse point cloud mapping based on selfsupervised point features is proposed.Relying on the advantages of both accuracy and efficiency of the proposed feature point detection algorithm,an efficient feature extraction module applied to the pose estimation framework is constructed to improve the performance of the pose estimation algorithm.Based on this,a multi-threaded sparse point cloud building framework is further designed by constructing a local-mapping thread and a loop-closing thread to output sparse point cloud maps in real time.Qualitative and quantitative experimental results on the KITTI dataset show that the proposed framework has reliable pose estimation accuracy.Compared with the pose estimation framework based on SIFT and Super Point algorithm,the m ATE of the proposed framework decreased by 57.3% and 8.2%,respectively.3.In order to further verify the performance of the proposed algorithm in the real-world environment,a miniaturized unmanned mobile platform for collecting image data is constructed.First,hardware devices such as a small unmanned vehicle,an embedded development board,and a camera are connected through serial communication.Then,the remote data communication of the unmanned platform is realized based on the topic communication of ROS.Finally,the realtime motion estimation of the unmanned platform is completed through the image data collected by the camera.The test results show that the pose estimation and mapping performance of the proposed algorithm in the real-world environment is reliable,which can meet the application requirements.In addition,a scalable and multi-functional feature point detection software is developed for testing and application of algorithms,which improves the efficiency of testing and application of algorithms through a simple visual operation interface. |
PDF Full Text Request