Efficient and Accurate Feature Extraction Using Local Density Detector

Feature detection is essential to a large number of vision-based applications. Among the approaches available, keypoint detection-based ones, e.g., SIFT, SURF, and ORB, are very popular. In particular, ORB stands out given its attractive balance of efficiency and efficacy, compared to other methods. However, a major drawback that affects the performance of ORB is the high density of keypoints it detects. In this work, a novel method namely local density enhanced ORB (ORBLD) is proposed. ORBLD mitigates ORB's weakness by adopting a local density detector to regulate the number of the keypoints. This approach achieves lower computational cost and reserves robustness under transformation and environmental changes. ORBLD is evaluated by setting up experiments with a self-driving related dataset, and the results show the reduction of 59.8% of keypoints mainly from redundant area, while the representative keypoints are reserved. ORBLD can facilitate the subsequent steps in feature extraction by optimizing keypoint selection and thus results in overall improved performance.