Research Of Object Detection Based On Contours

Object detection has been the significant issue in the field of computer vision, pattern recognition and machine learning for a long time. Although there have been proposed many methods for this research, these methods are still limited by some certain situations.In the wild, many factors influence the precision of object detection, such as, shadows,illumination changes, object deformations and occlusion. The object detection method based on contour features and the object detection method based on texture features can solve such problem to a certain extent. These two methods have their own characteristics.Compared with contour features, texture features including more information are more convenient to extracted. Contour features are more robust to illumination. Some partial contour fragments can provide the label information of objects. However, it always cost a large amount of computation to extract contour features and calculate similarities. Therefore, in this thesis, the main researches include the rapid feature extraction and similarity computation.The main contributions of the thesis are as follows:(1) A fast contour representation method is proposed. Firstly, the salient points are extracted by combining the chord-to-point distance accumulation(CPDA) and the elliptic plane curve model. Secondly, the least squares method is employed to fit the salient points of contour curves into hyperbolic curves which can fit symmetric and asymmetric curves.Experiments show that this method is effective and practical. On the one hand, we reduce the computational cost by creatively taking use of the CPDA during extracting salient points. On the other hand, since we consider the global property of curve fragments, the elliptic plane curve model is used for extracting salient points of target curves, which improves the robustness and practicability.(2) An effective object detection method based on salient segments is proposed.This method improves the precision of object detection through introducing the salient segment features into the framework of directional chamfer matching(DCM). At first, we represent the local structure information of shapes by CPDA so that we can accurately find the salient points of contours. Then, according to the links between salient points,we obtain the multi-directional chamfer distance maps. At last, we finish the matching detection based on DCM. Our method is evaluated on some datasets and achieves good results, which demonstrates the validity and the accuracy of our method.(3) A rapid object detection method using the two-stage backtracking is proposed.This method which is based on the way of index label backtracking builds the relationship of data transfer between two stages and implements the coarse-to-fine shape matching.In the coarse matching stage, we search objects on the distance map and obtain a possible object set. In the fine matching stage, firstly, we construct the description matrix of the shape contours according to the possible object set. Secondly, we build the multiobject optimization model based on the internal blocks of the description matrix and the dissimilarities between these internal blocks. Thirdly, we get the optimal solution of the multi-object model by Pareto method and obtain the final detection results. The experiments show that the computation cost are reduced greatly due to the coarse-to-fine shape matching and the detection speed is accelerated obviously.(4) A multi-class object detection method based on the index dictionary is proposed.This method votes for features on the target image according to the built index dictionary so that we can get the object locations rapidly. Firstly, the discrete multi-direction contour features are done the distance transformation respectively. Secondly, we take the quantity codes on the pixel scores of distance maps and their context features to build a feature index dictionary. Thirdly, the classes of features are decided according to indexes. And we vote to the centers of object classes of features and finish the object detection. The experiments show that our method can detect multi-class objects rapidly and simultaneously. Our method is evaluated on several datasets, which demonstrates that our method has a high generality.(5) A detection method of matching shapes based on local contour fragments of object shapes is proposed. This method designs the local contour fragment features, constructs the objective function according to template shapes and object contours. And we achieve the good results of matching shapes by dynamic programming method. At first,we extract local contour fragments according to the chord-to-point distance vectors and chord-to-point angle vector. Then, we construct the objective function which is used for finding the corresponding points of template shapes and employ the dynamic programming method to optimize this function. At last, we take use of the nonlinear optimization method to match the shapes between templates and objects. The experiments demonstrate that our method increases the accuracy and robust of the corresponding of two point sets between template shapes and object contours during the process of matching shapes.The thesis deeply studies object detection method based on contour feature, which includes some key issues about object detection, such as, shape representation and shape similarity matching, and achieves some research results.