| Summary: | 碩士 === 國立成功大學 === 電機工程學系 === 105 === Visual odometry (VO) or Simultaneous localization and mapping (SLAM) for camera poses estimation plays an important role in virtual object registration of augmented reality (AR) applications. One of the main challenges of the algorithm is to keep the tracking errors as small as possible for reducing the discrepancy between virtual and real objects which can be easily recognized in practical application. Moreover, it is hard to keep up with the camera frame rate while trying to maintain small error. Therefore, implementing a fast and robust visual odometry algorithm is pursued.
The proposed VO algorithm is based on iterative closest point (ICP) algorithm, which is a widely used registration algorithm but the quality of ICP estimation is easily affected by insufficient structural constraint and noise. Several methods have been proposed to sample points with enough transformation constraints; however heavy computation is necessary for knowing normal vectors of every pixels. In this thesis, we propose to sample points from near-edge regions, which can effectively increase constraints by increasing the possibility of sampling points from various objects in the scene. Besides, to decrease the influence of axial noise, we only sample points from near-edge regions inside the region of interest (ROI) which is extended from the near-end of the view, and then is adaptively extended or shrunk to ensure sufficient constraints while minimizing the influence of axial noise.
Secondly, we aim at reducing the complexity of matching stage in ICP. Based on multi-resolution scheme, an idea of adaptive searching area determination is proposed to reduce the redundant iterations as a smaller searching area is sufficient to search for true match in small camera motion. In addition, through using temporal correlation of ICP in steady state at the initial estimate for finding the closest point, number of search for each source points can be furthered reduced.
Finally, the proposed scheme is extended to the keyframe-based method where a keyframe is generated when the distance between the current frame and the last keyframe is large. Howerver, since the distance is unknown before estimation, a poor matching quality may occur, which results in large error; Hence, a strategy of efficiently changing keyframe and then performing re-estimation is proposed to increase the matching quality.
The proposed VO algorithm is evaluated on publicly available benchmark dataset. Compared with other VO algorithms, the proposed one exhibits competitive performance and achieves an average frame rate of 136 FPS using only a single CPU thread.
|
|---|
