Anomaly Event Detection Using Isolation Forest on Surveillance Videos

Detection of abnormal activities in surveillance videos is important for the security and safety of the public, especially in crowded public places. In this manuscript, a new approach is proposed to detect abnormal events using Local Binary Pattern (LBP) features coupled with the Isolation Forest algorithm. The LBP method is employed for the extraction of texture features from pre-processed video frames which captures important structural information. These extracted features are then used by the Isolation Forest algorithm, which is an unsupervised machine learning algorithm well-known for its capacity to spot outliers to differentiate between normal and abnormal events. It works by splitting the data points into sections randomly using feature selection and split values. This further results in an ensemble of trees with few and different anomalies being more prone to isolation and leads to shorter average paths within the tree structure, therefore effectively identifying anomalous cases. The performance of the proposed method is evaluated on two benchmark datasets- UCSD Ped1 and UCSD Ped2, for pedestrian anomaly detection in terms of various performance parameters like accuracy, precision, Recall, and F1-score. The maximum accuracy obtained is 88% for UCSD Ped 1 & 97% for UCSD Ped 2.