Vandalism Detection in Videos Using Convolutional Feature Extractor and LSTM Classifier

Actions or behaviors are typically categorized as being normal or abnormal. A technique used to find rare or unusual patterns, events, or data points that significantly differ from a dataset's expected or typical behavior is known as abnormal activity detection, sometimes referred to as Anomaly detection or outlier detection. Walking, running, and crossing oil pipelines are seen as regular actions in the context of this study, whereas excavating, drilling, and sawing on the pipeline are regarded as anomalous actions. A dataset designed exclusively for oil pipeline Vandalism detection (anomaly) and an algorithm that considers such specific acts which led to pipeline failure is not publicly available. Pipeline failure negatively impacts Humans and the Environment in no small measure. This paper aims to investigate and develop Activity Recognition Model to classify anomalous behavior for Pipeline Monitoring. A customized dataset was created and used to train and test the proposed Oil Pipeline Vandalism Detection Model. Moreso, the activities that led to the Vandalism and those that does not lead to it are thoroughly examined and recognized. The two main activity/action recognition processes/stages are feature extraction and classification. Feature extraction entails collecting significant feature vectors from a dataset (videos) that can capture the features of the behaviors, whereas classification looks for the presence or absence of specific patterns to classify a video dataset as normal or abnormal. For feature extraction, Google and Xception Fine Tuned Networks are utilized, and a comparison is done to determine the best Network. Experimental results reveal that Xception Net outperforms Google Network in terms of fast convergence, low Validation loss, and less elapsed time. Utilizing the feature vectors obtained from the Xception Network, Long Short-Term Memory (LSTM) is used to categorize the activity as normal or anomaly/vandalism. To reduce crimes like oil pipeline vandalism, which significantly impacts people's health, safety, the environment, and the economy, it is essential to have an automated anomaly detection system.