Machine Learning for Crowd Analysis and Classification

Abstract

Monitoring of crowd in public areas plays important role towards public safety and it is very hard to accomplish. Large population and Huge population and different harmful actions performed by humans leads to the importance of crowd analysis. Crowd management includes lots of challenges including analysis of crowd, identifying, monitoring and detecting anomalous activities. Conventual methods for crowd analysis were not very useful due to severe clutter and occlusions. In this paper we have proposed a machine learning-based classification approach for classification of crowd images. The work is caried out in two stages; crowd counting using density map estimation and crowd classification using machine learning. U-Net inspired model is used for crowd density map estimation and crowd counting. For crowd classification we have considered 3 categories of crowd images. The proposed algorithm of crowd image classification is divided to three stages; preprocessing, features extraction and classification. Median filtering is used for noise removal during preprocessing. Color and texture-based features are extracted for classification. Features are trained using various machine learning algorithms including LDA, CART, KNN, Naïve Bayes and SVM. Precision, recall, f measure and accuracy is used for evaluating the performance of machine learning algorithms. Maximum performance of 65% is achieved using HOG features with SVM with 30% test data and 70% data for training. The experimental results demonstrate that the proposed technique can be used for any type of crowd classification tasks