特征检测子是近年来计算机视觉领域的研究热点。在宽基线匹配、特定目标识别、图像及视频检索、机器人导航和数据挖掘等多个领域均得到了广泛应用。为总结特征检测子目前存在问题及指出可能的发展方向, 对比分析了当前普遍使用的特征检测子的鲁棒性与速度。首先综述了尺度不变特征检测子(SIFT)、快速稳健特征检测子(SURF)、二进制稳健尺度不变性特征检测子(BRISK)、定向的基于加速段检测子(ORB)、风式特征检测子(KAZE)以及快速风式特征检测子(Accelerated-KAZE), 然后按照特征检测步骤深入分析了各种影响检测子性能与速度的方法, 最后利用Mikolajczyk 05标准测试图集测定并分析了检测子的复现率与耗时实验。实验结果表明, 快速非线性尺度空间、基于加速段的特征检测子(FAST)与长距离对迭代法的综合性能较好。

Feature detectors are receiving increasing attention from computer vision research community, which have been widely utilized in a large number of applications, such as wide baseline matching, object recognition and categorization, image retrieval, visual search, robot localization and data mining.To discuss research challenges of investigation and directions for further research, analysis is made to the robustness and speed of current widely used feature detectors.Various feature detectors, including Scale Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), Binary Robust Invariant Scalable Keypoints (BRISK), ORB, KAZE, and Accelerated-KAZE are reviewed.Also, the algorithms affecting robustness and speed of feature detectors are investigated based on steps of feature detection.The Mikolajczyk 05 testing image sequences are used to determine and analyze repeatability and time cost.The experimental results show that fast nonlinear scale space, Features from Accelerated Segment Test (FAST) and long-distance pairs iteration have a more comprehensive performance.