Object detection and tracking are basic tasks in video surveillance and have been an active research area. Using a standard Gaussian Mixture Model (GMM) based method, nearby objects could be merged into a single foreground object. This causes difficulties in foreground segmentation, especially when objects in the foreground have similar in color, texture and shape. This paper proposes a novel method for segmenting merged objects into individual ones. First, an unsupervised co-training framework is proposed for the detection of foreground containing multiple vehicles. The co-training based approach is to simultaneously train two disparate classifiers based on independent features. One is a naive Bayes classifier based on scene context features, such as direction of motion and width of vehicles; the other is an appearance classifier based on Multi-block Local Binary Pattern (MB-LBP) features. Unlabeled examples which are confidently labeled by one classifier are added, with labels, to the training set of the other classifiers. The trained classifiers are then used to classify detected foregrounds into containing either a single vehicle or multiple vehicles. As the second step, foreground containing multiple vehicles is further segmented into individual vehicles by means of projection histogram analysis. Experimental results show the effectiveness and efficiency of the proposed framework.