Anomalous Human Action Recognition with Deep Learning Technique

Detecting and comprehending human actions poses a significant challenge, especially in the context of enhancing human security. This area of study has garnered attention recently, and machine learning-based methods have emerged as the forefront approach for identifying unusual activities. In this research, the main focus is on spotting abnormal activities in videos, utilizing the UCF crime database as the primary source of data. The proposed system uses a range of pre-processing methods on the video data from this collection. The system employs deep learning techniques with Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) algorithms. The frames extracted from the videos are fed into CNN, which processes and extracts important features. The LSTM module is responsible for categorizing the activities portrayed in the input video. By combining the strengths of CNN and LSTM, system successfully identifies six distinct anomalous activities: Abuse, Arson, Assault, Burglary, Fight and Robbery. For training and evaluating the system, divide the dataset into an 80% training set, a 10% testing set, and a 10% validation set, including cross-validation steps. The resulting model demonstrates an impressive accuracy rate of approximately 97.8%.