Crowd Density Estimation Based on Frequency Analysis

Numerous accidents from crowd stampedes have been recorded in human history. Therefore, the public has set high priority on the safety of public places. Surveillances systems aim to use artificial intelligence to address this problem, so that the crowd accident can be significantly reduced. We adopt a low cost camera to gather visual data and propose a cellular model for data interpretation. Based on the model, the motion status of the measured area can be represented as a dynamic state matrix, so the proposed method can save a lot of computing time. We adopted the Discrete Cosine Transformation to transform the motion status of the measured area into the frequency domain to recognize the frequency distribution. Then the feature values are extracted based on different frequency bands and distinct directional information to form a feature vector for training and classification. Finally, the Support Vector Machine is used to classify the feature vector into five classes of crowd density, with the results showing the proposed system is highly effective in crowd monitoring.