Research Of Homography Estimation Algorithm Based On Joint Feature And Photometry

Fast and accurate homography estimation is a fundamental task in many robotics applications,which can be formulated as an image registration problem.The traditional feature-based registration method has a large convergence area and relatively low accuracy,while the intensity-based method has higher accuracy and a relatively smaller convergence area.There is an urgent need for a combination of the two to achieve a better registration algorithm.Although the registration method based on deep learning has achieved good results,the number of parameters and the size of the model are large,which is not convenient for the application of embedded devices.Although some lightweight models greatly reduce the number of parameters,the registration accuracy is also reduced.This thesis conducts research based on the above questions,and the main research work is as follows:(1)Aiming at the problem of low accuracy and poor robustness of the image registration algorithm based on pixel intensity under the coexistence of large illumination changes and geometric interference,a Lie group space homography estimation algorithm based on pixel intensity is proposed.Joint geometric and photometric direct homography estimation in Lie group spaces.First,the joint geometry is co-parameterized with the non-Euclidean Lie group structure of the photometric transformation.Then,the second-order approximation strategy of ESM is used to optimize the geometric and photometric parameters.In order to further improve the efficiency,independent geometric and photometric parameter convergence criteria are used in the iterative optimization process,and the robust function is redesigned to deal with the unknown occlusion problem.Finally,the proposed algorithm is applied to real-time target tracking based on the robot operating system ROS,and it can still track accurately in lighting and occlusion environments.(2)Considering that the feature-based method has a large convergence area and relatively low accuracy,while the intensity-based method has complementary characteristics,an algorithm that combines the two characteristics is proposed.First,the Jacobian matrix in feature matching is reconstructed through the Lie algebra formula,realizing the same ESM optimization framework to better combine with the algorithm in Chapter 3.Then,the two classes of algorithms are unified into a single nonlinear optimization process,the same minimization method is applied,and adaptive weights are designed to combine the cost functions of the two.Finally,the proposed algorithm is applied to real-time target tracking based on the robot operating system ROS,which performs better in extremely challenging scenarios.(3)The parameters and size of the image registration model based on deep learning are too large,which is not conducive to the application of embedded terminal devices,and some lightweight models have poor accuracy.A compressed network model combining geometry and photometric loss is proposed,with only 1.428 M parameters and 5.54 MB model volume.Using group convolution,depthwise separable convolution and channel rearrangement operations to build a lightweight module,used in the redesigned regression network model,combined with self-attention mechanism,pooling and full connection for feature weighting,the model volume is larger than the existing The compression method is 51% smaller and more accurate in challenging environments such as lighting and large displacements.