Volume 18, No. 6, 2021

Suspicious Activity Detection In Crowd Videos Using Deep Learning Model

R.Thirumalaisamy , S.Kother Mohideen


Video surveillance systems have become a focal point of interest for both researchers and industries. With the emergence of advanced surveillance cameras equipped with powerful processing capabilities, the development of intelligent visual surveillance systems has become feasible. These systems are required to track crowded areas and detect abnormal activities, including unusual behavior and instances of violence. Due to the continuous nature of surveillance video, it is impractical for humans to manually track and analyze it. Therefore, the prime target of this work is to build a fully automated system that can analyze and identify suspicious activities occurring in crowded places. To achieve this, a deep learning model named Modified DenseNet201 is introduced. Input video sequences are preprocessed using filter to eliminate noise and are then converted into frames. Key frames are extracted, and the frames are augmented to expand the size of the database size. The segmentation task is addressed using U-Net. Following this, the modified Densenet201 is trained on the training data, enabling it to establish the relationship between input and output variables. The performance of the modified DenseNet201 is tested with test data. To validate the effectiveness of the developed system, two different datasets are utilized: an examination dataset and a violent/non-violent dataset. Experimental results demonstrate that the proposed system outperforms the other methods, producing excellent results on both datasets.

Pages: 9374-9387

Keywords: Abnormal behavior detection, deep learning, examination hall, random forest, and surveillance system

Full Text