Along the path followed by an autonomous vehicle, different positions are stored near the recorded route. These positions are being reported by the GPS device and transformed to the local coordinates to work easily with them. In a parallel way, the installed camera system is going to give to the application an image for each position that the car is passing over. For a given position, the system also looks for the nearest image in the database attending at different criteria.
In order to have the system working, a big database with images corresponding to the whole area where the prototype will be working has been generated. Each one of these images has information about the place where they were taken (horizontal and vertical coordinates and orientation), time of day, day of the year and other variables that are used in the selection of the most suitable image during the method execution. This image is compared with the image that is being obtained by the cameras installed on the vehicle.
The method is composed by three main steps:
1) Database Image Selection: The image is selected from the database attending to several criteria
2) Image pair registration: This is the main part of the algorithm and the most critical, as the good detection of the obstacles depends on a good aligmente while avoiding using so much computing time, which was the most limiting resource in the application.
- Feature Localization and Matching: A set of features are obtained from each image and later matched based on their descriptors. We evaluated the following detectors and descriptors:
+ Harris / Shi & Tomasi: In the first versions of the algorithm, the Shi & Tomasi implementation of Harris method was used for the detection of corners, as other descriptors like SIFT or SURF were too slow. The matching was performed by double-checking the optical flow between images parting from the detected features and going back to the original image.
+ SURF, PCA-SIFT, SURF: The performance of those three were compared, concluding that SURF gave a better ratio quality/speed. In the later implementations, a GPU-based implementation of this descriptor and a matcher has been used, allowing a reasonable performance for the method.
- Transform model estimation: Due to a big parallax in the images being processed, a local transformation must be performed. Piecewise Linear (PL), Thin-Plate Splines (TPS) and Weighted Mean (WM) methods were tested, obtaining the best results with the former.
- Image resampling and transformation: In this step, for each new position the appropriate value of the colour components is obtained by interpolation. The most common techniques, like the nearest neighbour, bilinear and bicubic interpolation have been tested, obtaining the best results with the bicubic interpolation.
3) Registered images comparison: By applying an Principal Component Analysis (PCA), both images are compared. Differences correspond to obstacles.
Creative Commons Attribution-ShareAlike 4.0
/
Search inscriptions
