为了获得能够很好地应用于远距离目标识别且计算快速的图像特征, 本文提出了一种结合尺度空间FAST(加速分割试验特征)角点检测器和SURF(加速鲁棒特征)描绘器的新特征算法。SURF算法利用了基于快速海森矩阵的关键点检测算法, 容易从图像中快速海森矩阵响应值较高但信息匮乏的边缘区域提取大量关键点, 进而导致大量的低独特性特征以及不可忽视的误匹配率; 同时, 其高斯滤波带来的图像模糊使得算法在远距离目标区域内检测到的关键点数量减少, 从而对远距离目标的识别造成困难。针对SURF算法的这些问题, 本文方法利用尺度空间FAST算法代替快速海森矩阵, 并利用具有良好的独特性的SURF描绘器。该方法能够有效地减少对上述类型的干扰性关键点的提取, 对远距离目标的关键点检测的性能相对于快速海森矩阵具有显著优势, 且其独特性优于同样使用FAST角点检测器的BRISK特征。实验结果表明, 对于带有光照变化、尺度变化和3D视角变化目标, 基于本文特征的识别算法的识别正确率高于基于SIFT、SURF和BRISK特征的识别算法; 本文特征适用于远程目标识别, 同时其计算速度达到了与SURF接近的水平。

In order to obtain image features that can be well exploited in long distance target recognition and are computationally efficient, an algorithm combining scale-space FAST (Features from Accelerated Segment Test) corner detector and SURF (Speeded Up Robust Features) descriptor, is proposed in this paper. The Fast-Hessian matrix based keypoint detector used SURF algorithm, is apt to extract numerous keypoints from non-informative edges with relatively high Fast-Hessian response, leading to considerable amounts of low-distinctive feature and consequently high rates of mismatch; with Gaussian filters employed, the possible amount of keypoints extracted with Fast-Hessian from regions of small targets is largely reduced due to image blur, which leaves difficulty for recognition of long distance targets. To address these problems, the proposed method uses a scale-space FAST corner detector in place of the Fast-Hessian detector, combining with SURF descriptor for its distinctiveness. The proposed method effectively eliminates the problem of extracting interfering keypoints along image edges, performing a significantly better detection of keypoints on long distance targets than Fast-Hessian, and generates features of better distinctiveness than BRISK features, which uses FAST as well. The experimental results indicate that the recognition algorithm based on the proposed features gives better performance against targets with change in scale, illumination and 3D viewpoint than that either based on SIFT, SURF or BRISK; the proposed feature can be well applied to long distance target recognition, while reaching a comparable computation speed to SURF.