Abstract: To cluster the appearance feature of the moving object plays an important role in object modeling, detecting and tracking applications. However, it is difficult to pre-set the cluster number for random data samples. In this paper, a novel framework of online determining the cluster number is proposed aiming at adaptively classifying the appearance features of vehicles. Iterative threshold values and RGB component detection fusion are utilized to automatically extract the object region from the background and restrain the background disturbance. The peak contour of the V（Value） color component histogram（under HSV space） in the object region is extracted. The sequential grayscale differences of adjacent candidate peaks and the residual peak energy are calculated and used to derive the cluster number. The sample intensity matrix of the object appearance is constructed with S（Saturation） and V（Value） components, and clustered with the obtained cluster number and K-means algorithm. Results from experiments show the excellent performance of the proposed algorithm in the class number adaptivity, the clustering effectiveness and the computational efficiency.

Key words: appearance feature, adaptive cluster number, peak contour, candidate peak, residual peak energy

摘要： 对运动目标外观特征聚类，在视频目标建模、检测和跟踪中具有重要应用。针对随机数据样本中类数难以事先设定问题，以交通车辆外观特征自适应聚类为对象，提出一种新的类数在线确定方法。采用迭代阈值和RGB分量目标检测融合，自动从背景中提取运动目标区域，以抑制背景干扰；在HSV空间中提取目标区域的V色彩分量直方图峰值轮廓；根据相邻候选峰的连续灰度差分和残余峰能量获得聚类数；利用S和V分量构成目标外观特征的样本强度矩阵，并联合所得聚类数和K-means算法实现样本强度聚类。实验结果表明，提出的方法具有类数目自适应性、聚类有效性和计算高效性。

关键词: 外观特征, 自适应类数, 峰值轮廓, 候选峰, 残余峰能量