3D model based object recognition is a classical problem in computer vision. The prior information provided by the 3D model can not only simplify the low level process-ing of image data and improve the robustness to noise and occlusions, but also make it possible to recover 3D information directly from image plane. In this dissertation, we focus on the 3D model based vehicle localization and recognition in traf.c scene surveillance. Novel algorithms are proposed and experimental results demonstrate the performance of these approaches. The major contributions of the dissertation include the following: 1. Camera calibration is the indispensable step of 3D model based object recogni-tion, which establishes the relations between the 3D model and image data. We propose a camera calibration method based on the appearance and motion infor-mation of objects in traf.c scene surveillance videos. With the camera height measured as the only user input, our method can completely recover both the in-trinsic and extrinsic parameters of the camera. Our practical camera calibration method greatly improves the applicability of 3D model based recognition. 2. Automatic object classi.cation in videos is the preprocessing step of 3D model based object recognition. We propose two object classi.cation methods to deal with perspective distortions of 2D image features. Both methods are based on online learning and avoid the acquisition of large database and manual labeling. Object classi.cation makes it possible to extract vehicle regions of interest from image plane so that we can initialize the 3D pose of vehicles and reduce the search space. 3. Fitness evaluation between the projection of 3D model and image data is the key problem in 3D model based object recognition. We propose a .tness evaluation method based on local gradient information of image data. The method avoids 2D geometric primitive extraction and distance calculation to show outstanding accuracy and ef.ciency. The good properties of the optimization surface offer convenience to the following .tness optimization. 4. The 3D model based vehicle localization, tracking and recognition can be seen as applications of .tness evaluation. We make use of gradient decent for .tness optimization to achieve 3D model based vehicle localization. Furthermore, we combine the .tness evaluation with the particle .lter based tracking framework, which shows good performance with robustness to noise. At last, we investigate the s...
修改评论