Scale-invariant and adaptive-search template matching for monocular visual odometry in low-textured environment
The most important task for any autonomous mobile vehicle is the reliable estimation of its position over time. Visual odometry (VO) is a localization technique that estimates the position of a robot using only the stream of images acquired from a camera. A monocular VO system that uses a single dow...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2016
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/70181/1/FK%202016%206%20IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The most important task for any autonomous mobile vehicle is the reliable estimation of its position over time. Visual odometry (VO) is a localization technique that estimates the position of a robot using only the stream of images acquired from a camera. A monocular VO system that uses a single downward-facing camera to estimate the relative position of a ground car-like vehicle at low-textured environments is presented in this thesis. In general, the main limitations of existing VO systems are related to computational cost and light and imaging conditions such as sunlight, shadows, image blur, and image scale variations.Fluctuations in camera height from the ground when driving on an uneven terrain can lead to variations in image scale and, in turn, affect the accuracy of vehicle position estimation. Therefore, a new technique and algorithm were developed to resolve the image scale uncertainty. This technique marks the image frames by using two laser points as independent reference points. It can also estimate and adjust image scale variations by monitoring the variations in distance between the two reference laser points. The proposed technique improves the accuracy of camera motion estimation to less than 1% error and dispenses with the necessity for camera re-calibration when the number of passengers and the load in the vehicle change. It can likewise replace the usage of sensors, such as a laser range finders or inertial measurement units,to measure the variations in camera height.Normalized cross-correlation template matching was utilized to estimate the pixel displacement between the image frames by computing the degree of similarity between them. This method is one of the most effective methods for
template matching. However, it incurs high computational cost because its underlying mechanism depends on a series of multiplication operations.Therefore, an adaptive-search template-matching technique based on vehicle acceleration was developed to reduce the correlation computational cost,increase the allowable vehicle traveling speed and reduce the probability of template false-matching. Size of template and search area were determined and calculated to reach a trade-off between the performance and computational cost of template matching. This developed technique sped up the correlation process with more than 87% reduction in computational cost compared to the traditional full-search correlation. The factors that affect the maximum permissible vehicle driving speed were also determined and the related equations were derived. The maximum allowable vehicle speed for the developed VO is up to 6.3 m/s.In short, the developed VO system, as well as the proposed algorithms and techniques, were successfully implemented, tested, and validated using real time kinematic GPS (RTK-GPS) with 2 cm positioning accuracy. Several indoor and outdoor experiments were conducted, and the results displayed high efficiency for the suggested techniques. Hence, the developed techniques and algorithms have high potential to be implemented in various commercial mobile robotic applications, which utilize VO for improved accuracy, efficiency, and cost effectiveness. |
|---|