There is a glut of multimedia content shared on various public communication sites like YouTube, Facebook, etc. every day. This multimodal content contains information on various issues such as terrorism, child pornography, product reviews and much more. Censoring on-line data that promotes unethical activities globally cannot be done manually. This has generated the need for a tool which can automatically determine the central idea expressed in the video and also categorize it into different emotion labels. When classifying audio video content into different class labels, audio-visual features are highly interdependent. Naive Bayes classification model fails to classify them correctly. In this paper, we propose and have implemented, in the form of a web interface, a novel approach for performing efficient multimodal sentiment analysis of online videos using machine learning algorithm - Perplexed Bayes classification technique, which takes care of the inter-dependency among the features. The interface displays the emotion label - happy, sad or neutral depending on the content of the video, the degree of emotion and the keywords related to the uploaded video. Further, Perplexed Bayes classifier has been proved to give 28% better results than Naive Bayes classifier.