Hybrid Deep Learning Convolution Neural Network Scheme for Enhanced Human Activity Recognition System in Security Surveillance

Abstract

Nigeria continues to experience a high security breach with the attendant loss of lives and
properties worth in billions especially in the North-East as ravaged by insurgency,
kidnapping and other forms of crime that have overwhelmed law enforcement agencies.
Technology have been explored to help curb security threats and improve surveillance. This
study proposes an enhanced human activity recognition system with hybrid deep learning
fuzzy convolutional neural network. We utilize dynamic agile development mode using the
python IDE for training and testing the machine learning models. The CNN-Fuzzy model
was trained with 3,500 dataset extracted from 7 classes of the UCF Crime dataset which was
further split into 70:30 ratio for training and test purposes. The model after 30 epoch with
training and validation accuracy of 0.9954 and 0.9954 yielded a prediction accuracy of
97.14%, Recall of 97.14%, F1-Score of 96.83% and precision of 114.2 outperforming the
existing system making it an efficient tool for security surveillance to mitigate security
breaches and generate early warning signals for security agencies.