Automatic classification of criminal activities for security surveillance by keyframes detection and advanced inception techniques

The exponential growth of video surveillance systems in public and private spaces has heightened the demand for effective and automated abnormal behavior detection mechanisms. Traditional surveillance systems rely heavily on human operators, making them prone to fatigue, subjective errors, and inefficiency. Recent advances in Deep Learning (DL) techniques have emerged as transformative tools to address these limitations due to their superior ability to analyze complex spatio-temporal patterns and extract high-level features from vast amounts of video data. The Convolution Neural Network (CNN) is widely used as a standard model for image processing because they are designed to learn features from images focusing on regions, extraction details like textures, edges and shapes without manual feature engineering. The model ability lies under the combines multi-scale factorized convolutions with residual connections. In this research study, the latest version of Inception which is a Google based model and is enhanced version of CNN and has been applied on standard datasets for abnormal behavior detection. The main novelty lies in the introduction of key frame extraction method which is used in efficient models by recognizing and selecting key frames only and thus reducing data size. The comprehensive empirical analysis-based results reveal that the proposed Inceptionv4 shows the highest results of 95% as compared to several standard deep learning algorithms and pre-trained models.