Abnormal Behavior Detection: A Comparative Study of Machine Learning Algorithms Using Feature Extraction and a Fully Labeled Dataset

Although the number of surveillance cameras in public spaces like streets, banks, parks, shopping malls and others is rising considerably due to low costs of implementation and quick access to technology, the monitoring capability has not increased proportionally. Detecting abnormal behaviors using computer vision and pattern recognition is a long standing challenge. After the research of previous work solutions, we decided to fully label, on a segment level, a dataset with abnormalities, used a generic 3D convolutional neural network to extract feature vectors of each segment and trained a Multilayer Perceptron to do the classification of normal and abnormal behaviors. Our contribution consists, firstly of a fully labeled dataset that is composed of 16853 videos where 9676 videos are labeled as normal and 7177 are labeled as abnormal. Secondly, by the use of the labeled dataset on our proposal, our method outperformed the results of our baseline research with an Area Under the Curve (AUC) of0.863. Finally, we compared our results with other classifiers to demonstrate that the use of a segment-labeled dataset definitely improves the results of the classifiers tested.