Image features for Visual Teach-and-Repeat Navigation in Changing Environments


We evaluate the robustness of feature extractors to naturally-occurring environment changes. Our evaluation is based on the ability to register images from similar locations, which is crucial for vision-based teach-and-repeat navigation. Images of outdoor scenes across seasons are particularly difficult to register. Our evaluation uses five datasets gathered in three different countries. The first two datasets contain images gathered on a monthly basis at several locations during one year. The three remaining datasets contain image sequences captured by moving vehicles in different seasons. To evaluate the image features,

15
00:01:05,501 –>00:01:08,000
each image pair from every location was registered and the result was compared to manually-established ground truth. Before the registration, the locations of binary comparison pairs that form the core of the BRIEF feature descriptor were trained by a simple evolutionary method using the images from the Planetarium dataset. We call the trained BRIEF descriptor `GRIEF’. The evaluation of each feature extractor was based on its success rate in registering all

26
00:01:40,501 –>00:01:43,000
image pairs from all datasets’ locations. When using a low number of keypoints the Superpixel Grid (SpG) detector combined with the CNN descriptor outperformed the other evaluated features. However, the Superpixel Grid method did not provide as many keypoints as other detectors and ultimately, the STAR detector and GRIEF descriptor outperformed SpG/CNN on 2 datasets out of 5. Since the performance of the image features is influenced by both the detector and descriptor phases, we evaluated several detector/descriptor combinations. The experiments confirmed that both SpG/CNN and STAR/GRIEF features achieved low registration error rates, making them suitable for long-term teach-and-repeat vision-based navigation in naturally-changing outdoor environments. While the SpG/CNN feature achieved lower registration error rates in semi-urban areas, the STAR/GRIEF performed slightly better on datasets with significant

51
00:02:48,001 –>00:02:50,500
appearance changes caused by seasonal variations of foliage. Compared to the SpG/CNN, the STAR/GRIEF is computationally inexpensive, which makes it suitable even for resource-constrained robots. However, the STAR-GRIEF is tailored specifically to teach-and-repeat navigation and thus, SpG-CNN will perform better in more general scenarios. The complete evaluation framework including the datasets is available at http://purl.org/robotics/grief

Leave a Reply

Your email address will not be published. Required fields are marked *