Evaluation of Feature Detectors and Descriptors for Motion Detection from Aerial Videos

In tasks of motion detection from aerial videos, feature-based image registration is an essential step to compensate ego motion of airborne vehicle between consecutive frames. This paper presents the first performance evaluation of feature detectors and descriptors for image alignment and frame difference to detect pixels with motion from aerial videos. To this end, we design two criteria, namely Position Error Rate(PER) and Correct Match Rate(CMR), to characterize the registration accuracy and frame difference success rate, respectively. To generate the pixel-wise registration ground-truth, we employ sophisticated block-matching method, which is then checked and corrected manually by control-points-based alignment method. Based on the proposed metrics and ground-truth registration parameters, five detectors (Harris, FAST, SUSAN, DoG, and SUSAN_M) and four descriptors (Intensity, BRIEF, HOG and SIFT) are examined. We test detector-descriptor combinations in typical visual light aerial videos and infrared aerial videos. We find that detector plays a more important role in both registration accuracy and efficiency than descriptor does, thus should receive more attention in the area of motion detection from aerial videos. For detectors, DoG performs well in most videos but has the lowest efficiency, and SUSAN_M achieves good performance balance between registration accuracy and efficiency. We also reveal that currently widely used detectors should be tailored to moving object detection tasks in future research on the aspects of feature spatial layout, removing features on moving targets, feature number control, as well as computational efficiency.