Suspicious Object Detection in Environments with Tear Gas

In a digital security surveillance system, detecting suspicious objects is paramount to identifying criminals and thereby reducing the crime rate. At present most of the security surveillance systems are based on automated CCTV cameras. Usually, it is highly challenging to detect objects in environments covered with smoke, water, and dark clouds. One of the main intentions of security surveillance is detecting harmful objects such as guns, knives, axes, and iron rods. Digital surveillance in a situation where the police have used tear gas and water is even more challenging since the background could contain many artefacts. This paper aims to develop a model to detect suspicious objects in footage of protests or riots where the environment encompasses tear gas. The proposed method uses a YOLOv4 object detection framework along with Roboflow. All the fully connected dense layers of the YOLOv4 model were trained with a newly created dataset with different images captured in environments where tear gas is present and containing dangerous objects. Also, the developed dataset uses several Bags of Freebies to improve the detection rate. The model achieved a mean average precision (mAP) of 82% for the test dataset when trained for 105 epochs. Overall, the proposed method could solve problems in suspicious object detection in public security surveillance under challenging environments.