Abstract:Efficient complete intelligent video surveillance system has become the urgent needs in today's society.For the crowded scenes involving multiple complex motion modalities,an unsupervised automatic detection group algorithm in the crowd is proposed,which firstly extracts Kanade-Lucas-Toma (KLT) points from the foreground fields based on Gaussian Mixture Models (GMMs) background subtraction strategy.Then for each foreground-keypoint,velocity direction filtering operator,motion correlations filter operator is applied to move behavior inconsistent neighboring points,and non-neighbors are filtered based on path-based similarity.Thus,all foreground-keypoints are merged and split each other to achieve group detection.This method does not need single pedestrian segmentation,training sample and prior information;meanwhile,the number of neighbors is self-tuning according to the nearest distance neighboring keypoints in foreground fields.To test the performance of this algorithm,videos with multiple complex motion modalities are used in experiments.Experimental results show that the proposed algorithm is more efficient and more robust in complex crowded motion scenes.