quotDetecting Multiple Moving Objects and Interpreting their motion Pattern in Crowded Environmentquot

Abstract

With the reduction in the cost of cameras and increase in the crime rate, Video newlineSurveillance has become commonplace. Cameras can now be found in Airports, newlineRailway stations, Malls, Banks and almost at every public places. With 24 x 7 video newlinerecording coming from lakhs of cameras, it becomes imperative that means be newlinedeveloped for automatic analysis of the video. In the field of Computer Vision, newlineautomatic analysis of surveillance video to detect anomalous behaviour has been an newlineactive area of research. newlineThe first step in this direction is detection of moving non-rigid objects by blob newlineanalysis. In the current literature, methods have been proposed for detecting objects in newlineun-crowded videos. Detecting moving objects in crowed environments like railway newlinestations, marathons etc is a challenging task, more so because of the small size and newlineunpredictable behaviour of the non-rigid objects. In this work, we have proposed new newlineMotion Detection Algorithm (MDA) using Fuzzy Neural Network for detecting newlinemoving objects in crowded and densely crowded environments. The method is based newlineon adaptive and dynamic template matching. The features (volume and perimeter) newlinefrom the detected moving objects are used to train the neural network. The output newlinefrom the neural network is used by Lagrangian Support Vector Machine (LSVM) to newlineclassify rigid and non-rigid objects in the video frames. newlineOnce moving objects have been detected, the next task is to track their motion. For newlinetracking non-rigid bodies in crowded environments, a hybrid tracking model is newlineproposed. In densely crowded environments, generally only heads of humans are newlinevisible. The heads are tracked using an objective function which is a weighted sum of newlinecolour histogram and texture. The trajectory of the moving objects (heads) is found newlineusing Zero Stopping Constraint based on two models Structure Similarity Model newlineand Time Series Model. The advantage of using the proposed hybrid model is that it newlinegives a rich representation of the trajectory and the computation time is considerably newlinereduced. newlineThe