Target Detection Based On Deformable Part Model

As a very active research field in computer vision,pattern recognition and machine learning,target detection algorithms are closely related to people's lives,and have a subtle influence on human production and life style.In recent years,the target detection algorithm has been developing rapidly,but it is still a challenging problem.Based on the study of predecessor's research,this thesis studies the target detection based on the deformable component model algorithm.The object detection algorithm based on deformable component model is an excellent algorithm for object detection.In the PASCAL VOC competition over the years,the deformable part model and its related algorithms have achieved excellent results.Has become the mainstream of the common target detection algorithm.Since it was put forward,a lot of improved algorithms have been proposed,which greatly improved the accuracy and efficiency of the algorithm.In this thesis,based on previous research work,after the in-depth study of the deformable part model,we improved the algorithm of cascade deformable part model from three aspects.The main work of this thesis is as follows:First,n the model training feature extraction part,this thesis chooses 14-dimension HOG eigenvector for model training,and reduces the model training time by reducing the eigenvector dimension of the algorithm.Secondly,the feature of the image to be detected is further processed by the SLIC super-pixel image segmentation technique to realize the separation of the foreground and the background of the HOG feature,so as to reduce the influence of the background on the target object.Check the situation,and improve the detection accuracy of the algorithm.Thirdly,after the convolution is completed,the cascade deformable part model does not consider the hypothetical location and the adjacent area information to the target hypothesis trimming stage,ignoring the connection between the two.In this thesis,we will consider the position and neighborhood information to take into account,speed up the cutting speed,and further improve the rate of hypothetical cutting,thus enhancing the efficiency of detection.Finally,the improved algorithm is applied to the PASCAL VOC data set and the real image to verify the effectiveness of the algorithm.The experimental results show that the improved detection algorithm does reduce the training time to a certain extent,The detection accuracy,speed up the detection efficiency of the algorithm.