Local visual homing by warping of two-dimensional images
Robotics and Autonomous Systems
Three 2D-warping schemes for visual robot navigation
Autonomous Robots
| Hi-index | 0.00 |
This thesis concerns visual homing in robots and proposes solutions judged by both performance and biological plausibility. Visual homing is the ability to return to a goal from a nearby location using vision. Biological plausibility is defined as the likelihood that a proposed algorithm mirrors the internal mechanism of an animal solving the same problem. The high effectiveness of insect homing motivates interest in biological plausibility. The approach to biological plausibility is through retinotopic processing---a characteristic of natural visual systems. Experiments are conducted on a database of panoramic images collected from a variety of indoor environments. A review of existing methods reveals Franz et al.'s [26] warping method as the de facto standard for robot visual homing. The warping method is compared with two other methods which adhere to the constraints of retinotopic processing. Its performance is found to be superior and is therefore used for subsequent performance comparisons. The first new homing method proposed is an extension of Möller et al.'s [76] neural snapshot model. It is found to be applicable to real-world images, but is not competitive with the warping method. Two other new methods are then proposed based on a novel scale invariant image descriptor. One of these is found to perform nearly as well as the warping method. Further, its scale invariance property is found not to be crucial to its success. Thus paving the way for simpler methods. The first such simpler method, based on block matching, is found to outperform the warping method. Analysis reveals that this method succeeds by pairing low-frequency image components and that increasing the number of correspondences is beneficial. Two even simpler variants based on matching pixel intensity and the gradient of intensity were also found to perform quite well. The success of these variants prompts an investigation into differential-based techniques. Two methods are proposed and both found to perform competitively. These methods are found to be successful because of the presence of regions of small feature shifts in the image near the foci of expansion and contraction. The differential methods satisfy both the performance and biological plausibility criteria.