An Invariant Backward Feature Analysis of Model-Based Malicious Activity Monitoring for Efficient Video Surveillance Using Deep Learning
K. Lokesh, M. Baskar · Engineering Technology & Applied Science Research · 2025
The issue of malicious activity monitoring in video surveillance has been extensively studied, and several methods have been developed to address it. These approaches typically rely on attributes such as shapes, objects, textures, and sketches; however, their accuracy remains limited. To overcome these shortcomings, this paper presents an effective Convolutional Neural Network (CNN)-based malicious activity monitoring approach. The proposed technique detects harmful behavior by leveraging the invariant properties of drawings and their spatial positions across multiple preceding frames. To enhance input quality, Layer-Based Feature Normalization (LBFN) is applied to recorded video frames, removing noise and improving clarity. Feature segmentation is then performed using the Value-Oriented Segmentation (VOS) algorithm. The model maintains features extracted from the previous k frames and incorporates them to extract features from the current frame. Convolution and max-pooling layers are employed to convolve and normalize the extracted features. At the output layer, Sequential Position Support (SPS) and Sequential Sketch Support (SSS) are calculated using a variety of activity-related characteristics retained by the model and are iteratively evaluated across frame and feature sequences that the model has produced. Based on this process, the technique computes Malicious Activity Support (MAS) scores for different activity classes and assigns higher values to the most probable class. The proposed Invariant Backward Feature Analysis Model Convolutional Neural Network (IBFAM-CNN) achieves an accuracy of 97% in malicious activity monitoring and video surveillance.