The purpose of this project is to discover the most appropriate combination of key-point detectors, descriptors, and matching techniques for the collision avoidance system relying upon the Time-to-Collision (TTC) metric calculated based on the sequence of images provided by a mono camera mounted on top of a vehicle based on the trade-off between the number of key-points discovered, the number of matches between the consecutive images and the processing time required to produce one estimation of TTC.

The default OpenCV implementation parameters were used for each entity (detectors, descriptors, matchers, selectors). The measurements were taken on a laptop having the GPU: NVIDIA GeForce GTX 1060 Max-Q 6GiB; CPU: Intel Core i7-8750H; RAM: 16 GiB.

To see all the details, open the image above in a new tab and zoom.

To see all the details, open the image above in a new tab and zoom.

To see all the details, open the image above in a new tab and zoom. The values near each box present the mean values.

Below is a list of keypoint detectors sorted in descending order based on the mean number of keypoints identified among 10 images. The analysis is based on the keypoint data presented in the section Keypoints Statistics.

Below is a list of TOP-5 combinations of detector+descriptor sorted in descending order based on the mean number of identified matches among 9 pairs of 10 consecutive images. The analysis is based on the matches data presented in the section Matches Statistics.

Below is a list of TOP-5 detectors sorted in descending order based on the mean time required to identify keypoints in one image. The analysis is based on the timings data presented in the section Timings Statistics.

Below is a list of TOP-5 descriptors sorted in descending order based on the mean time required to compute a set of descriptors for all the keypoints in one image. The analysis is based on the timings data presented in the section Timings Statistics.

Assuming that the number of keypoints, number of matches, and timings ratings contribute equally, the TOP-3 detector-descriptor combinations are chosen.

The choice between BRISK+BRIEF and FAST+BRIEF should be based on the hardware available and accuracy requirements. From visual observations it is clear that both combinations produce some false matches on each pair of images, that is why it is probably better to choose FAST+BRIEF combination for collision avoidance system, given the significant gap between the FAST and BRISK keypoint detection times.

The winner is FAST detector + BRIEF descriptor combination.