ORB and AKAZE Feature Tracking in Mobile Markerless Augmented Reality: An Experimental Study at 10 Lux

Abstract

Markerless Augmented Reality (AR) relies on reliable visual feature detection and matching to maintain image correspondence between consecutive frames. Under low-illumination conditions, reduced image contrast and increased noise can affect the availability and geometric consistency of local features, which may influence mobile AR tracking. This study experimentally evaluates the feature-tracking characteristics of ORB and AKAZE under a nominal 10-Lux illumination condition using a mobile-camera markerless AR evaluation scenario. A total of 300 frames from one video sequence were processed, generating 299 consecutive frame pairs. Five complementary metrics were evaluated: keypoint count, Mean Processing Time (MPT), frame-to-frame reprojection RMSE, Feature Matching Rate (FMR), and Inlier Ratio. ORB produced substantially more keypoints (335.79 ± 31.39/frame) and lower processing time (15.21 ± 1.52 ms), whereas AKAZE produced lower reprojection RMSE (0.057 ± 0.094 px vs. 0.153 ± 0.098 px) and higher FMR (95.53 ± 11.27% vs. 86.76 ± 18.06%). Inlier Ratio was very high and nearly identical for both algorithms. Frame-level Wilcoxon Signed-Rank analysis indicated systematic paired differences for all five metrics (p < 0.001). Because consecutive frame pairs were temporally related, these results were interpreted as within-sequence evidence rather than evidence from independent experimental replications. A temporal-block sensitivity analysis preserved the direction of the main observed differences, while no systematic block-level difference was observed in Inlier Ratio. The study provides a controlled multidimensional characterization of ORB and AKAZE under the tested nominal 10-Lux condition.