Keypoints Based Object Detection With Generalized Hough Transform

From car detection, video retrieval to Microsoft’s Xbox360, applications based on computer vision have brought revolutionary changes to human life. Object detection is the basis of those applications, and is an important research technique of pattern analysis and computer vision. It not only needs to consider the accuracy of detection, but also needs to meet the requirements of practical application. However, the methods of object detection with the high accuracy often base on highly complex algorithms, and real-time performance may degrade accordingly, which is not conducive to practical application. What’s more, traditional Hough transform can only be used for simple shape detection, which is very slow. Considering the requirements, this thesis has done lots of work for object detection based on generalized Hough transform and keypoints based feature extraction. The work consists of the following two parts:(1) Feature extraction based on keypoints. Image feature is the basis for object detection. Good feature would describe the information of the object well, which could improve the accuracy of detection. After analysis of image feature such as color, texture, shape and spatial relationships, this thesis give descriptions of the commonly used features including SIFT, HOG and BRISK. The research on keypoints-based feature focused on the extraction of keypoints, the proposed FAIR-SURF and the parameter optimization of DAISY. FAIR-SURF first generated the simulated image set through the selected affine parameters, and employed the improved SURF for feature extraction. The proposed algorithm combined the full affine invariant of ASIFT with the fast computation of SURF. The experimental results of image matching have verified the effectiveness of the algorithm. DAISY describes more generally for the non-extreme points, and is robust to light, rotation and so on. It uses the fast convolution map of images in multiple directions to extract features quickly. This thesis optimized the parameters of DAISY with two rings and eight directions. Experimental results have shown that the optimized DAISY could greatly reduce the complexity in the premise of ensuring the performance, which is more conducive to the practical application.(2) Object detection based on the generalized Hough transform. Based on the analysis of traditional object detection technology, this thesis studied the implicit shape model, the mechanism of discriminative Hough voting, and built a new object detection framework with the generalized Hough transform based on keypoints. The traditional Hough voting has too many candidate points, thus it is slow. To solve this problem, the proposed framework preprocessed the images before the extraction of keypoints-based feature, which effectively reduced the number of candidate points. Meanwhile, the framework used the optimized DAISY features as low-level features of objects, which improved the detection speed and was more robust to rotation. In order to verify the framework, this paper carried out experiments of hand detection, and employed the look-up-table-based Gaussian mixture model for pre-processing. Experimental results have shown that the proposed framework has high detecting rate and gains good real-time performance. This framework could meet the demand of common object detection, especially for the partially occluded object.