Anomaly Event Detection and Classification Using Deep Learning Model and Deep Double Q-learning Network
B. Vidhya, V. Malathy, Piyush Kumar Pareek, Haider Mohmmed Alabdeli, Javvadi Lakshmi Prasanna · 2024
Abnormal event detection is pivotal in surveillance systems, which is utilized for identify or detecting unusual events or behavior of humans in sequence of video frames. The existing solutions for event classification face difficulties in distinguishing abnormal events and normal events from input videos. To solve this problem and to improve classification process a Deep Reinforcement Learning (DRL) model is proposed in this paper for precise identification of normal and anomalous events. The surveillance videos from UCSD dataset are converted into video frames for feature extraction. Resnet-152 is utilized for extracting important features and anomaly events are detected by the Faster-RCNN. Finally, the detected events are classified as normal or anomalous by the integration of a Double Q-learning with Deep Neural Network (DDQN). The results of the presented DDQN is evaluated by the performance metrics attained Precision of $\mathbf{9 9. 2 \%}$, Recall of $\mathbf{9 9. 1 \%}$, Accuracy of $99.2 \%$ and F1-score of $\mathbf{9 9 \%}$ is superior than previously implemented Deep Learning (DL) models are Deep - Q Network (DQN), Bilinear Convolutional Neural Network (BCNN), CNN-Support Vector Machine (CNN-SVM), DL- Social Force Model and Fully CNN (FCNN).