使用频谱词包的接力跟踪目标交接方法

Method for relay tracked target handover using the spectrum bag of words

针对接力跟踪中目标跨摄像机交接困难问题,提出了使用频谱词包的目标交接方法.首先对第1摄像机的两幅或多幅目标图像分块提取傅里叶特征;再对特征集进行K均值聚类,选择其中紧密类的质心特征作为正常单词,另引入一个异常单词,与正常单词一起共同构成词包;最后用词包分别描述两摄像机的目标,以欧氏距离度量相似性.实验表明,该方法的目标识别性能和鲁棒性比目前几种常见的基于纹理的方法有一定的提高.

Aiming at the difficulty of the relay--tracked target handover across cameras, a method using the spectrum Bag Of Words（BOW） is proposed. Firstly, 2 or more target images from the first camera are divided into blocks. The Fourier features of the blocks are extracted and then clustered by Kmeans. The averaged vectors of the compact clusters are used as the normal words of the BOW, while an abnormal vector is used as the abnormal word. The targets from the first and second cameras are described in BOW words. The Euclidean distance between the BOW descriptions is used as the target similarity metrics. Experiments show that the proposed method has some definite improvement in target recognition and robustness compared with several common methods based on the texture.