Object Recognition Research From Point Cloud In 3D Scenes

Object recognition in cluttered scenes is a fundamental research area in computer vision. In the last few decades,2D object recognition has been extensively investigated and currently been a relatively mature research area. Compared to 2D images,3D point cloud provide more geometric information, as result, the pose of an object estimated from 3D point cloud is more accurate than that from 2D images. Moreover, the rapid development of affordable 3D information collection devices (e.g., Kinect) makes point cloud data more accessible. All these advantages make 3D object recognition an active research topic. It has numerous applications, such as robotics, scene understanding, virtual reality, human-computer interaction and laser remote sensing measurement.Existing 3D recognition methods can be divided into two broad categories:local feature based methods and global feature based methods. The local feature based methods extract only local descriptors around specific keypoints. They are more robust to clutter and occlusion. However, feature point matching can lead to many mismatches. On the other hand, the global feature based methods recognize the object as a whole, so the recognition speed is very fast. While pose estimation becomes another technical difficulty.To solve these problems mentioned above, we proposed to apply graph matching to feature correspondences and take full advantage of global and local methods. Based on PCL, the proposed approaches achieved preferable recognition performance. The main work of this paper includes:1. A novel approach for 3D object recognition based on graph matching. The key step of local feature based methods is to find consistent correspondences between two sets of features. Geometric Consistency algorithm is the classic solution, while we obtained local optimization because of dealing with only one correspondence each iteration. Regarding the feature correspondences as graph matching issues, we adopted an approximate graph matching strategy to solve the object recognition problem from the perspective of global. The evaluation showed that our method worked well.2. A 3D object recognition method which fuses global and local features. The object recognition system aiming at scene understanding has a high requirement to real time and precise pose, the proposed method used a hybrid technique based on global and local features. We adopted a global feature to recognize the object, whereas, a local feature was used to estimate the position of the object in the real world. The experimental results revealed that the method performed well.3. A recognition prototype system which facilitates point cloud data research. We designed the software platform system based on Qt. It implemented plenty of functions such as visualization of point cloud and feature histogram, preprocessing, model training, feature calculation, segmentation and object recognition.