Abnormal Activities Detection along Oil Pipeline Using Deep Learning

Abnormal activities like oil pipeline vandalism need to be identified promptly. Manual surveillance systems for oil pipelines use ground team surveys, while CCTV Cameras are employed in semi-automated surveillance to detect those abnormal behaviours. Oil pipeline failure resulting from vandalism has detrimental effects on both humans and the environment. Despite the availability of the current technologies, escalating incidences of vandalism occur, prompting the necessity for computerized monitoring techniques. Computerized solutions that use deep learning networks require an enormous quantity of information for their implementation. The popular UCF Crime dataset is meant to detect generic vandalism and other anomalies of a similar or divergent nature. Hence, a dataset explicitly designed to complement such a model and assist in pipeline monitoring is needed. This work aims to investigate and develop a behaviour recognition model and a new dataset named Vandalism Detection Dataset 2024(VDD 24) for detecting and classifying abnormal behaviours along oil pipelines. A Modified pre-trained ResNet18 is used for feature extraction, and a Bi-directional long-short-term memory (Bi-LSTM) is employed to detect and categorize those human actions as normal or abnormal (vandalism). Digging, sawing, hammering, and stone impact are regarded abnormal, while activities such as walking, running, and cycling are defined as normal. Despite the similarity in the abnormal actions in the dataset, the model was able to detect and classify the anomaly. Experimental results reveal that our model’s performance on VDD 24 is significant, with an accuracy of 82.5%. The model is further validated on the UCF-Crime dataset with an impressive performance.