Model matching based object recognition and pose estimation in a sequence of depth image
A recognition of a relatively big and almost stationary object, such as a refrigerator and an air conditioner is necessary because this object can be crucial a global stable feature of Simultaneous Localization and Map building(SLAM) in indoor environment.
In this paper, we propose a novel method to recognize and estimate a pose of the big object in a sequence of scenes captured from a stereo camera. In order to use consecutive scenes effectively, the particle filtering method is used. The particles representing the possible pose of object are scattered into the environment and then the probability of each particle is calculated by matching with the 3D lines in the environment. The final pose of object can be determined based on the probabilities and the convergence rates of the particles.
The proposed approach successfully works in texture-less environment, although the stereo camera cannot provide enough depth data in that environment and the experimental results show the feasibility of incremental object recognition based on the particle filtering method and the application to SLAM.