Crowd Anomaly Detection via Multiscale Features and Zero-Shot Learning in Public Surveillance

Improving the precision of video surveillance systems in identifying unusual activities is essential for maintaining public safety. The efficacy of traditional approaches is weakened by their frequent struggles with high rates of false positives and negatives. In this work, we present an advanced method to tackle these issues by classifying behavior into normal and deviant classifications using the Avenue dataset. Frame extraction is the first step in our technique, which is then followed by grayscale conversion, filtering, binarization, inverse transformation, and region shrinkage to lessen occlusion impacts. Aggregated Channel Features (ACF) detect human presence, and fuzzy c-means clustering is used to confirm the results further. Our sophisticated feature extraction methods include particle gradient motion analysis, Histogram of Oriented Gradients (HOG), and Harris corner recognition. Next, zero-shot learning is used to classify these characteristics and Particle Swarm Optimization (PSO) is employed for optimization. With a mean accuracy of 93.0%, our suggested approach considerably lowers false positives and negatives, improving public safety by precisely and consistently detecting anomalies.