关键词: k-means聚类, 视觉显著性, 运动目标检测, SURF特征, 帧差法

Abstract:

The traditional target detection algorithm based on feature point matching has low target recognition rate and high false detection rate because of inaccurate matching of feature points and target contour discontinuity. We introduce the improved spectral residual algorithm and k-means clustering algorithm to solve those problems, and propose a moving target detection algorithm based on spectral residual algorithm and clustering algorithm. The method is divided into two parts. Firstly, we extract the speed up robust features (SURF) from every two frames and complete image registration. Then, we use the spectral subtraction algorithm to extract the saliency feature from frame difference results in order to remove noise and false targets caused by inaccurate matching. Secondly, the improved k-means clustering algorithm is introduced to cluster the discontinuous contour curves to obtain a complete target after morphological processing. Experiments show that the target recognition rate of this new algorithm is 90.61% and the false detection rate is 21.23%, which are better than the traditional moving object detection algorithm based on SURF whose corresponding results are 66.60% and 31.91% respectively  and  the moving object detection algorithm based on improved oriented FAST and rotated BRIEF (ORB) whose corresponding results are 87.57% and 26.80% respectively. Although the average running time of the algorithm is 18 frames/s, it can meet the requirement of video fluency. This algorithm therefore can be used as an effective moving target detection algorithm in dynamic background.

Key words: k-means clustering algorithm, visual saliency, moving target detection, SURF, frame difference method