Enhanced Anomaly Detection in Surveillance Footage

With the proliferation of surveillance cameras, managing large volumes of visual data is a prime concern for security. This work proposes an automatic anomaly detection system based on feature extraction using MobileNet and Bi-LSTM-based temporal analysis that focuses on road accidents and violent behaviors. With training on different types of anomalies such as theft and accidents, the model demonstrates 97% accuracy, 95% precision, and 95% recall, which outperforms existing methods. Implemented using TensorFlow and Keras, the framework boosts the surveillance system's realtime anomaly detection ability. Future work involves combining with self-supervised learning and optimization for edge computing.