Multi-objects recognition using unsupervised learning and classification

The objective of this paper is to develop a real-time unsupervised learning method to detect multi-objects. Variance and gradient variance, as main texture feature, are compressed based on PCA (Principal Component Analysis) to get initial classifications of clusters via K-means algorithm in the image frame. The cluster-kernel of each class, the nucleus as we defined, is figured out through shifting the sampling area in each cluster. Based on the nucleus, the policy of cell expansion as we designed is operated to merge different classes into one object which turns the work from feature level into object level. All potential objects and cells of the objects are detected and labeled in the frame. The descriptors of each object are employed as hypothesis for the next frames object-classification. Our results demonstrate that it is possible and fast to recognize multi-objects without and training model or labeled data. The process of learning could be operated with initial clustering and automatic updating with information of the new classification.