Research On Procedural Geometric Model Fitting Theory And Its Applications In Pattern Recognition

The task of geometric model fitting is to model the given data by finding the ground-truth models(e.g.,lines,circles,characters,or buildings)that best explain the given data point set(e.g.,images or laser scanning point clouds).It has a wide range of applications,such as curve reconstruction in reverse engineering and computer vision,and the reconstruction of three-dimensional(3D)surfaces of buildings in complex urban scenes.Most existing model fitting methods were proposed to fit classical models.However,there are many complex data that cannot be explained by classical models.The procedural model fitting studied in this paper is a generalization of the classical model fitting.In theory,it can model arbitrarily complex data.Therefore,compared to the classical model fitting,the procedural model fitting will undoubtedly have more profound theoretical research significance and a broader practical application value.However,the procedural model fitting has the following four main difficulties:First,how to design the similarity estimator to ensure that the ground-truth model has the unique largest similarity,so as to ensure that the optimization algorithm can find the ground-truth model.Second,how to design the optimization algorithm to quickly find the ground-truth model.Third,how to write a probability program to facilitate fast and correct fittng;Fourth,how to properly process the original data to facilitate the fitting.In general,the first and second difficulties are common for all applications,while the third and fourth difficulties are related to specific applications.Therefore,this thesis studies first the procedural model fitting in theory,that is,proposes new solutions to address the first and second difficulties.The thesis then explores the following three applications of procedural model fitting:image-based handwritten character recongnition,vehicle-borne mobile laser scanning point cloud-based highway curve and building surface 3D reconstruction.Specifically,the main work and innovations of this thesis are as follows:(1)To address the first difficulty,the thesis proposes a new similarity estimator to estimate the similarity between geometric model and data.The proposed estimator is based on the error from model to data,while most existing estimators are based on the error from data to model.Experimental results demonstrate that,the proposed estimator is more strit and more robust than the state-of-the-art estimators.(2)To address the second difficulty,the thesis proposes a coarse-to-fine model dividing strategy to reject dissimilar models in advance to accelerate the optimization process.Experiments show that the current popular optimization algorithms(Markov chain Monte Carlo and cuckoo search)can be accelerated by approximately three times using the early rejection strategy.Moreover,the thesis proposes a deep reinforcement learning based multi-model fitting method.Experimental results indicate that the proposed method is tens of times more efficient than cuckoo search algorithm in terms of the numbers of fitting iterations.(3)The thesis applies the procedural model fitting method to achieve pattern recognition on images and laser scanning point clouds.The first application is to reconstruct character models on eighteen variants of imperfect MNIST image datasets to achieve few-shot pattern recognition.In the 5-shot recognition,our method outperforms the state-of-the-art method(George et al.2017 Science)[22]on thirteen variants of the imperfect data.In particular,for one of the data corrupted by grid lines,our method obtains a high accuracy of 65%,whereas the state-of-the-art method only obtains an accuracy of 30%.The second application is to reconstruct 3D highway curves from vehicle-borne mobile laser scanning point clouds.To address the fourth difficulty,a road marking extraction method based on intensity variance is proposed.An absolute accuracy of 20 cm has been achieved in estimation of circle radii based on the virtual scan dataset.A comparative study also showed that our road marking detection algorithm is more effective than the algorithm based on intensity gradient[137].The third application is to reconstruct 3D procedural building suface models from vehicle-borne mobile laser scanning point clouds.Experimental results show that the proposed method can handle models represented by context-sensitive shape grammars with continuous parameters,while most existing methods assume that the models are represented by context-free grammars with discrete parameters.In summary,this thesis thoroughly studies procedural model fitting and proposes some new solutions for some challenging issues.The thesis also explores several applications in pattern recognition and achieves state-of-the-art results.The major limitation of the proposed procedural model fitting method is its slow computational speed,making it inapplicable for the tasks that require fast computational speed.Further improvements such as acceleration of the proposed method are needed to be studied in future.