Abstract: A small targets classification algorithm in video surveillance is presented.First,a set of reliable,independent,and discriminative features are extracted according to the maximal mutual information.Then,the classifier with multi class support vector machines of DDAG(decision directed acyclic graph)is given.The training of the classifier is divided into two steps.The first step is to train and get the baseline classifier using scene independent features.The second is to train the base classifier furtherly using the scene dependable and scene independent features,so that the classification precision can be improved effectively.Experimental results show that the classifier not only is good at the classification precision but also has excellent adaptability to new scene.