Research On Object Detection Based On Hough Transform In Complex Scene

Object detection is one of the most fundamental and classical problem in the field of computer vision.Its core process is to identify categories,location and size of the object in test image according to specified detection algorithm.Although object detection has made many achievements,the performances of detection decline when object appearance changes in different complex practical scenes due to many challenging factors,such as occlusion,poses,viewpoints,illumination variation,nonrigid deformation.With the increasing demand of applications,how to achieve robust detection performances under complex environment is more and more important and urgent.The works are based on Hough Transform termed HT.According to characteristics of object in various scenes,the detection models are constructed by combining HT with some image representations to achieve robust detection.The main works of the thesis are concentrated on the followings:(1)The part of the works is built on HT and proposes a method of go game image detection.In previous works,many methods used HT to detect lines of go board.But at the end of the game,the grid lines will be covered more and more pieces.The above method greatly reduces the performance of grid lines.Considering go game scene,sequence is applied to go image detection.In general,the board and the camera don't move during the game.Since HT has a good result at the beginning of competition,HT is adopted to detect unoccupied board and then locate four corners of go board.Based on the above,the position of four cormers is fixed during the entire game.That is to say,four corners of unoccupied board are used to replace four corners of go board in each moving of go game.(2)It proposes simple geometric transform model termed SGTM.According to the locations of four corners in(1),SGTM can accurately find 19 rows,19 columns and 361 corners of go board.Experimental results show that the method has better performance on multi-view and occlusion.(3)HT of circle detection is applied to identify pieces.Many methods address the problem with color space.Only color space leads to the instability of piece recognition because of uneven light.In our method,pieces are identified by HT fusion color space.Firstly,HT for circle is adopted to locate piece.Then,k-means is used to cluster pieces into 3 clusters by color space:the black,the white and the empty.The method is robust to nonuniform illumination(4)Motivated by weighted Hough voting detection framework,an extension named improved weighted Hough voting detection method is proposed to solve multi-view car detection problem.Firstly,features computed by Convolutional Neural Network are made to obtain high-level local image patches.Then,Hough Forest is defined to generate visual words by clustering local image patches.Finally,combining discriminative weights for multi-view objects is used to determine object center.Experimental results demonstrate that CNN-based local image patches with powerful representations.Compared with weighted Hough voting detection framework,the proposed method shows better performance.