Abstract
In this paper, we propose a single SVM-based algorithm to classify moving objects inside videos and hence extract semantics features for further multimedia processing and content analysis. While standard SVM is a binary classifier and complicated procedures are often required to turn it into a multi-classifier, we introduce a new technique to map the output of a standard SVM directly into posterior probabilities of the moving objects via Sigmoid function. We further add a post-filtering framework to improve its performances of moving object classification by using a weighted mean filter to smooth the classification results. Extensive experiments are carried out and their results demonstrate that the proposed SVM-based algorithm can effectively classify a range of moving objects.