Research on video description and human activity recognition has been dramatically improving study finding on visual monitoring. Monitoring activities in the examination is a yet unsolved problem in which the students can perform various activities in an Exam room. Such activities can be monitored automatically through an automated surveillance system. We use squeeze net and VGG16 as deep learning constructs for deep feature extraction. These features are then fused serially to form a single feature set. The entropy and ant colony optimization (ACO) based feature selection approaches are applied separately on the acquire feature subsets having qualities of both filter and wrapper-based approaches. The separately selected features are then ensemble to obtain a powerful features subset. SVM based classifiers are finally applied for prediction. From the Exam activities detection dataset, the classification algorithm precisely labels the student activities into abnormal and normal classes. The results depict that the suggested framework for activity recognition in Exam is very effective with acceptable accuracy. The framework will help to analyze student activity in exams, to improve the examination system.