针对大多数特征提取算法忽略时间因素对识别精度影响这一问题,提出了一种新的时空局部特征提取方法.首先采用Harris算子提取关键点并估算出该关键点处的尺度,并使用无迹卡尔曼滤波器对关键点处的位置进行跟踪,获取不同时刻下的关键点簇；再采用小波系数描述关键点簇的特征区域、采用SSD衡量关键点簇上相邻两时刻特征向量的相似度,并保留随时间推移SSD值变化缓慢的关键点簇；最后使用高斯统计模型对这些关键点簇的特征向量进行统计建模,获取时空局部特征.实验结果表明,文中方法的目标识别精度高于基于SIFT的目标识别精度约10％.%Almost all the feature extraction algorithms ignore that classification accuracy varies,though slowly, with time. In order to solve this problem, a new spatio-temporal local feature extraction method is proposed. Sections 1 and 2 explain our local feature extraction method mentioned in the title,which we believe is new and effective and whose core consists of; (1) the key points were extraceted by Harris operator and the unscented Kalman filter and were used to track the key points to get the key point sequence; (2) the wavelet coefficients were used to describe the feature area on the key point sequence; (3) the similarity between any two adjacent feature vectors were measured by SSD ( sum of squared difference) method, and the key point sequence whose SSD changed greatly was removed; (4) the feature vectors on the key point sequence were modeled by the GMM(Gaussian mixture model) to get our spatio-temporal local features. Finally, the experimental results, given in Fig. 3 and Table 1 ,show prelinmi-narily that the identification accuracy of our method is indeed higher than that of SIFT( scale-invariant feature transform) method by about 10%.
展开▼
