An Efficient Motion Based Group Level Activity Recognition for Intelligent Video Surveillance

Human crowd analysis is a critical topic in video surveillance and a demanding process because of the non-rigid shapes of the crowds. Due to occlusions, tracking and analysing every single person in a crowd is a difficult operation. As a result, rather than following every individual in the crowd, the crowd can be treated as a single entity. Due to notable spatio-temporal properties, the crowd's behaviour can be differentiated by motion patterns. Optical flows are initially estimated and then used as a clue to group human crowds into clusters using the suggested adjacency matrix-based clustering algorithm (AMC). First, video is converted into frames and noise is removed in the stage of pre-processing. Second, foreground blob is extracted in the stage of background subtraction. Third, in the stage of motion tracking, velocity, position and direction of the motion are obtained by estimating the optical flows. Fourth, clustering is done using AMC algorithm in the stage of clustering. Fifth, features are extracted by using Harris algorithm along with centroid, orientation etc. and finally, the human crowd is detected and labelled using multi-class Support Vector Machine (SVM) algorithm. Obtained experimental results on PETS and UMN video datasets with a recognition rate of 95.6% and 95.1% respectively show the effectiveness of the system.