Design and implementation of Ensemble learning and swarm Intelligence algorithm for event Recognition in video surveillance

Abstract

In recent years, event recognition has attracted growing interest newlinefrom academia and industry. Recognizing events in surveillance videos is newlinestill quite challenging, largely due to the tremendous intra-class variations of newlineevents caused by visual appearance differences, target motion variations, newlineviewpoint change and temporal variability. The low image resolution, object newlineocclusion and illumination change in surveillance videos further aggregate newlinethe event recognition challenges. Mitigate these challenges, various works in newlineevent recognition turns the focus to context. The existing context approaches newlinefor event recognition either utilize context directly as feature inputs to newlineclassifiers like Support Vector Machines or incorporate context through newlinetraditional probabilistic graphical models like Bayesian Network, Markov newlineRandom Field or Latent Topic Model. It does not provide the satisfactory newlinerecognition accuracy. newlineAn Improved Hybridized Deep Structured Model for accurate newlinevideo event recognition is introduced. The feature level, semantic level, as newlinewell as the prior level contexts are introduced. Two types of context features newlineincluding the appearance context feature and the interaction context feature at newlinethe feature level are proposed. These feature level contexts exploit the newlinecontextual neighborhood of event instead of the target. In this model, the newlinesemantic level context captures the interactions among the entities of an newlineevent and the prior level context includes the scene priming and dynamic newlinecuring. These extracted interaction context features are grouped by using newlineImproved K-means algorithm newline