Anomaly Detection in Videos for Chain Snatching Using Meta-Learning

Anomaly detection in videos from CCTV cameras can be an important strategy for crime analysis and prevention. The main focus of our work is on detecting the crime of chain snatching from videos captured in India. Due to the absence of a training set of similar Indian videos, it is challenging to design a classifier for this crime. Hence a technique called Model Agnostic Meta-Learning(MAML) is used to train a network on the well-known UCF crime dataset for detection of chain-snatching in a dataset custom built by us. MAML is further developed to result in a method called Sampling-based Meta-Learning Anomaly Detection (SMLAD). With this, the characteristics of MAML are used automatically to classify chain-snatching as an anomaly and obtain best accuracy and AUC scores of 86% and 84%, respectively. Thus the proposed work demonstrates the efficacy of MAML to correctly classify chain-snatching which constitutes completely unseen data, as a crime-related anomaly.