Human Detection In Thermal Images Using A Vision Transformer

With technological progress, monitoring systems concerning security issues have gained a significant role. Among such systems, the majority of them are dedicated to human detection in indoor environments. Currently, the most popular monitoring solutions use cameras adapted to operate in the visible spectrum of light; however, thermal cameras deploying infrared radiation can be utilised to increase their effectiveness. This type of imaging can be exploited to detect people using appropriate algorithms. To develop such algorithms, substantial datasets comprising adequate images are required. Most databases accessible on the Internet include human depictions in urban environments. Therefore, to create a detection method for indoor utilisation, the first step was focused on developing an extensive database of thermal human depictions in an indoor environment. Then, a vision transformer was designed and evaluated with a comparison with state-of-the-art convolutional network models. The results showed that the proposed vision transformer architecture accomplished a high metric of mean average precision (0.601) with the lowest execution time (24 milliseconds).