Adaptive AI-driven deep neuro-fuzzy model and superpixel segmentation for complex anomaly detection
Abstract
Public safety and crowd control face major issues due to the world's expanding population, varied social relationships, and crowded urban environments. Such settings frequently make people feel uneasy, so strong surveillance systems are required to address safety issues. Traditional monitoring systems are ineffective for identifying anomalies in real time and require a lot of work and effort. A unique multi-stage architecture for automated monitoring and anomalous behavior identification is proposed in this paper to address these problems. After preprocessing and improving video frames using adaptive median filtering, superpixel-based segmentation is used to sharpen object boundaries and reduce background noise. A thorough feature extraction process is then carried out, using the following methods: dense optical flow for motion analysis, centroid-based radial shape descriptors for silhouettes, PCA-enhanced Gradient Location and Orientation Histogram (GLOH) for texture, and stochastic triangulated shape descriptors for structural geometry. While a neuro-fuzzy system is used to fuse and improve extracted features for dimensionality reduction and robust selection, Faster R-CNN is used for human and occlusion detection. To simulate geographical linkages and temporal dependencies for anomaly detection, a ConvLSTM network is shown. The suggested system exhibits robustness against complex motion patterns, fluctuating population densities, and occlusion. On the XD-Violence and UCF-Crime datasets, however, detection efficiencies of 88.05% and 94.34% are attained individually.