A computer vision based approach for stalking detection using CNN-LSTM hybrid model

Abstract

The next level of revolution toward a better world could involve combining human security with machine intelligence. In recent years, stalking in public areas has become a pervasive issue, and women are disproportionately affected. In order to solve the problem, we want to design a model that can identify public-space stalking. There have been several study papers and publications written on the topic of stalking. However, most of them relied on spatial co-occurrence for the detection of suspicious actions and face recognition, which does not adequately address the problem. Using a hybrid mix of CNN and LSTM, we explain in our study a model for determining the presence of a stalker situation utilizing a dataset of video footage. The proposed model was evaluated using two approaches: one using manual feature extraction and the other using dynamic feature extraction. The manual feature extraction approach was evaluated with three distinct machine learning classifiers (SVM,KNN, and Random Forest), whereas the dynamic feature extraction method was examined with two different CNN models (VGG16 and ResNet50) and a CNNLSTM hybrid model. The CNN-LSTM hybrid model has the highest accuracy of any of these models, at 89%. Experiment results indicate that the CNN-LSTM hybrid model detects a stalking scenario with a spatio-temporal advantage and provides a better classification result than other models.