3D Object Tracking And Reconstruction Based On Kinect

To reconstruct the moving object can help robots perceive the moving object in the scene. The output model can also be the data source of 3D printing. Different from the reconstruction of static object, we need segment the moving object from the scene first then to reconstruct the object with its own pointcloud. It will challenges more because of lacking points. To reconstruct it accurately, we divide the system into tracking and reconstruction.The first step is to denoise the depth image collected by Kinect. In the tracking stage, we focus on the way to make the tracking more robust with the depth image. So we propose a tracking method fusing multi-features. In the reconstructing stage, we use ICP to register the pointclouds. We propose a registering method with 3 frames instead of two. Moreover, we add a fault-tolerant module and register the pointclouds partly. Finally we get the 3D model of the object as the output. The main contribution made by us are as follows:1. We propose a tracking method with multi-features. This method based on particle filtering combines with RGB color histogram, outline of depth image feature and LBP texture feature. We judge the reliability of each model, then get the highest weight particle as the final result. As the tests show, our method can adapt to the change of light and object's rotation.2. We propose a registering method with 3 frames instead of two since two frames registration can lead to inconsistent match and transformation which affects the accuracy of registration. Considering the consistency, the result of the registration is more accurate.3. We propose a part registration with fault-tolerant module. As a consequence of tracking failure and registering failure, it will cause the overall registration error. So we add a fault-tolerant module to estimate the wrong registration and re-register from this frame then go to the next round registration. As the tests show, our method can adapt to the wrong registration and re-register again. The output is the 3D model.