Human Anomalous Activity detection with CNN-LSTM approach

Abstract

Human action recognition in digital media is a challenging task. Its processing has become popular recently due to increasing demand in fields related to human security. Machine learning based classification and recognition is the prominent approach for anomalous activity detection. The present work focuses on video anomalous activity detection. UCF crime database is used in the present work. Preprocessing operations are performed on videos from the dataset. The present system employs cascaded approach using Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM). The frames are applied to CNN as an input. CNN process the input and extract the prominent features. Convolution and pooling are the main operations in CNN. LSTM is used for the classification of the activities performed in the input video. Cascading of CNN and LSTM recognizes the six anomalous activities identified as Abuse, Arson, Assault, Burglary, Fight, and Robbery. 80% dataset is used for training, 10% for testing and 10% for validation and cross validation. The developed system attains approximately 99% accuracy and is a robust model.

URN:NBN:sciencein.jist.2024.v12.704