Research On 3D Point Cloud Object Recognition And Hyperparameter Optimization Based On Stacked Broad Learning System

With the continuous development of deep learning,models for point cloud and multi-view-based 3D point cloud recognition have been proposed,achieving significant advancements.Existing methods combine point cloud and multi-view data through complex deep learning structures to learn unified 3D shape descriptors,increasing the computational complexity of multimodal feature fusion.Therefore,in this thesis,a three-dimensional point cloud object recognition algorithm based on the fusion of multi-view and point cloud data is proposed.Stacked broad learning system has proven to be effective for one-dimensional data.Hence,this algorithm performs fast three-dimensional point cloud object recognition by directly using concatenated multimodal features as inputs to the SBLS.To address the issue of setting hyperparameters for the stacked broad learning system,this thesis introduces a hyperparameter optimization algorithm based on a multi-dimensional adaptive regression spline,which adaptively adjusts the parameter range.This algorithm determines the optimal hyperparameters for the broad learning blocks in the stacked broad learning system.The main contributions of this thesis are as follows:1.The proposal of a data-driven adaptive interval hyperparameter optimization algorithm.By converting the optimization of expensive black-box functions into the solution of a multi-dimensional adaptive regression spline regression model,this algorithm reduces the computational resource consumption of evolutionary algorithms.The effectiveness of the algorithm is validated by comparing it with existing algorithms on the NORB classification dataset and ten UCI regression datasets.The experimental results demonstrate that the proposed algorithm achieves superior hyperparameters.2.The proposal of a three-dimensional point cloud object recognition algorithm based on multi-view and point cloud data.Firstly,the multi-view features of the multiview branch are extracted using Res Net.Then,these features are fused with the point cloud features extracted by Point MLP and input into the stacked broad learning system.The effectiveness of this algorithm is validated by comparing it with state-of-the-art methods on the public point cloud datasets Model Net40 and Scan Object NN.PV-SBLS achieves 95.5% accuracy on the Model Net40 test set,surpassing existing algorithms.PV-SBLS achieves 87.0% accuracy on the Scan Object NN test set,improving 1.6%over Point MLP accuracy.Furthermore,a comparison of time efficiency with existing multi-view and point cloud-based methods demonstrates that the proposed method effectively increases training efficiency.3.Based on the research of three-dimensional point cloud object recognition algorithms,a system for object recognition based on multi-view and point cloud data is developed.This system is developed using the open-source programming library Point Cloud Library(PCL)and the C++ graphical user interface library QT.It implements functions such as point cloud data reading and visualization,point cloud denoising,multi-view generation,and three-dimensional point cloud object recognition.Through this system,researchers can conduct related research on three-dimensional point cloud object recognition and use its provided basic testing tools for validation and optimization of related techniques.