Database
The database used in this project consists of 39,853 images containing 91,731 annotated objects. Figure 1 highlights the distribution of annotated object types, with 68,309 being people, 11,617 being weapons, and 11,805 being helmets. It can be accessed through the following link: dataset
As previously stated, the database consists of images that prioritize the presence of people, helmets and weapons. The qualities of the images vary according to the device that was capturing them, but follow the pattern of being captured by surveillance cameras, this being a mandatory characteristic for the image to be in the base.
Figure 1: Distribution of annotated object classes
Figure 2: Annotated scenarios that display the distinctions between the data obtained
One consequence of looking for mostly human handled items is to have a large magnitude of entities classified as people. The base contains people in various positions and angles because the cameras are arranged in numerous positions and the people in focus are performing different actions in the same context.
Figure 3: Images containing people
In addition to the focus on individuals, the dataset also aims to strengthen the detection model of clusters, in other words, even if the individuals are clustered in such a way that they are not detected as unique entities, the value as positive reinforcement for the clustering model still makes the picture not negative for the project as a whole.
Figure 4: Images containing crowded people
The scenarios in which the weapons appear do not require other entities to be present, being detected in several situations. Furthermore, because of the variety of quality of the surveillance cameras and the way they are displayed, coloured or not, there is a difficulty in detecting weapons in certain angles that make them look like black rectangles generating confusion for the model. However, this problem can be alleviated by annotating the weapons together from the user's hands, giving more room for detection.
Figure 6: Images containing people with guns
The weapons detected by the model vary in size and colour, but since most of the time there is reinforcement that they are being held in one hand, it is easier to detect them without using specific classes for them.
Figure 7: Images featuring guns and their different shapes
The assembly of the helmet dataset isn't much different from the guns dataset, however, what makes it easier to compose the detection of entities as helmets are their shape.
Figure 8: Images containing helmets
Even in a state where the helmet is open, its delimitation is still easier than a weapon that is secured in the fact that it is being handled, which can cause problems on a large scale.
Figure 9: Images featuring helmets