Scalable Anomaly Detection Framework in Video Surveillance Using Keyframe Extraction and Machine Learning Algorithms

Vijay A. Kotkar and V. Sucharita

Since from last decade automatic anomaly detection in video surveillance gained significant attention from researchers to monitor and secure human lives from the treats. The range of fundamental support from the video helps to reduce this computation cost and time before applying the pre-processing, features extraction, and classification steps. This paper proposed the novel framework for anomalous event detection in order to achieve scalability and efficiency with minimum computational requirements. The proposed framework based on key steps like keyframe extraction, visual and motion features extraction, and classification. The framework is simple, but yet effective for improving the detection accuracy with less computational efforts. For keyframe extraction, we apply the dynamic threshold-based approach which may vary for each video sequence. The threshold is computed by estimating the difference among the consecutive frames using a histogram called Histogram of Difference (HoD). HoD extracts the keyframes and discards the other frames without loss of any event-specific frame. For feature extraction, we designed a hybrid technique using the Histogram of Gradient (HoG) and interest point extraction algorithms. The classifiers such as ANN, SVM, and KNN used for the class at the end. The simulation outcomes show that the suggested model improves the performances as compared to existing methods.

Volume 12 | Issue 7

Pages: 395-408

DOI: 10.5373/JARDCS/V12I7/20202020