Machine Learning-Enhanced ORB Matching Using EfficientPS for Error Reduction

Abstract

The study comes up with a new way to improve the accuracy of image matching in binocular vision systems, especially those that are mounted on vehicles. It combines machine learning with the ORB (Oriented FAST and Rotated BRIEF) image-matching algorithms. Standard ORB matching frequently encounters mismatches in complex and repetitive environments. To minimize false positives in matches, our strategy utilizes the EfficientPS (Efficient Panoptic Segmentations) algorithm, a panoramic segmentation technique that uses machine learning in conjunction with ORB. The procedure begins with the EfficientPS approach, which delivers fine-grained and efficient segmentation of images, assigning semantic category labels and unique identifiers to each pixel. The ORB feature point matching process is refined using semantic data to filter out mismatches between foreground objects and the background effectively. This machine-learning-augmented method significantly decreases the frequency of erroneous matches in intricate settings. Empirical findings from the KITTI dataset demonstrate that in non-targeted environments, the accuracy of our proposed method (0.978) is marginally less than that of LoFTR (0.983). Still, it surpasses other methods when utilizing 50 ORB parameters. In more intricate situations, such as multi-target scenarios with an increased number of ORB parameters (200), our method maintains a high level of accuracy (0.883), outperforming the conventional ORB (0.732) and rivaling the performance of DL-BDLMR (0.790) and ORB-MFD-FPMC (0.835). Our method’s processing time is competitive and slightly higher than the standard ORB, but it improves accuracy. In scenarios without targets and with single targets, our method’s processing time (0.195 seconds and 0.211 seconds, respectively) is greater than that of ORB and ORB-MFD-FPMC. Yet, it is significantly lower than that of LoFTR. In multi-target scenarios, our method’s processing time (0.226 seconds) is considerably better than LoFTR’s (0.445 seconds), effectively balancing processing speed and accuracy. We highlight the efficacy of incorporating machine learning to enhance the robustness of image-matching algorithms in dynamic and complex environments.