A Framework for Identifying Theft Detection Using Multiple-Instance Learning

Closed-circuit televisions are widely used devices to prevent theft and dangerous events. Video surveillance for unusual activity detection is done by human supervision. This work involves constructing a convolutional neural network to identify normal and abnormal activities in an input video. In this model, the data is preprocessed and fed into a fully connected three-dimensional convolutional neural network model for feature extraction. The extracted features are sent into a three-layered, fully connected network for processing. The ranking loss function is calculated using multiple-instance learning to predict the anomaly score. The anomaly score obtained for a given segment is used to map the instance to the anomaly class. This model outputs a result that provides insights into the accuracy rate concerning the number of epochs run by the model. Based on this observation, the model is trained with normal and abnormal videos to identify the anomaly.