Study Of Image Texture-less Object Detection Algorithm Based On Template Matching

The technology of object detection in digital image is widely used in many fields such as Robot system,Augmented reality,3D Reconstruction.It is an important research topic in the field of computer vision.In many practical industrial applications,a lot of objects have litter texture information,which are mainly described by its outline and shape,and the surface of objects has a large number of homogeneous regions.Such objects are called texture-less objects.Due to the lack of stable and discriminative feature points of texture-less objects,the visual features cannot be extracted steadily and reliably,posing a huge challenge to the traditional texture-based object recognition methods.Therefore,the rapid and accurate texture-less object detection has important basic meaning in both theoretical research and engineering application.The main works of this thesis include three aspects: study of object candidate location detection algorithm,study of texture-less object recognition algorithm,the acceleration of object detection algorithm based on GPU and CUDA programming model.After analysis the advantages and disadvantages of the current representative of texture-less object detection algorithms,this thesis proposes a fast approach to texture-less object detection based on template matching.At its core,the proposed method demonstrates a two-stage template-based procedure.The first stage predicates object candidate locations,scale and rotation with the fast-speed.The second stage effectively screens object candidate locations to determine the final pose of the object.Finally,the algorithm is accelerated by GPU parallel computing.The proposed algorithm in this thesis is divided into two stages: object candidate locations detection and object detection.The first stage of the method describes the representation of the template features based on the orientation compressed map,and then uses a similarity measure process to estimate candidate locations and corresponding scale and rotation information.The application of orientation compression map greatly reduces the complexity of template matching.In the second stage of the algorithm,a new template feature is built based on the discriminative region weight.To eliminate the false candidate locations,the second stage computes the similarity of each candidate location by discriminative region weight.The application of discriminative region weight has better performance for similar targets between different categories and effectively reduces the occurrence of false detection.Finally,the proposed method is compared with other representative algorithms in different image datasets.Experimental results show that the proposed algorithm performance has a fast recognition speed when it achieves the best object detection rate.In the aspect of image acceleration with GPU,based on the analysis of the parallel image processing technology with CUDA programming model,including the edge extraction algorithm,object detection algorithm.After analyzing the parallelism of the proposed algorithm,to take full advantage of the advantages of the GPU multi-core,the template feature data are mapped to the three-dimensional texture memory space by using the three-dimensional data volume.Finally,the thesis implemented the texture-less object detection algorithm on CPU + GPU platform.After the parallel acceleration,for the 1280×960 images,experimental results show that the proposed algorithm runs 20 times faster than the CPU.