Video streams from panoramic cameras represent a powerful tool for automated surveillance systems, but naïve implementations typically require very intensive computational loads for applying deep learning models for automated detection and tracking of objects of interest, since these models require relatively high resolution to reliably perform object detection. In this paper, we report a host of improvements to our previous state-of-the-art software system to reliably detect and track objects in video streams from panoramic cameras, resulting in an increase in the processing framerate in a Jetson TX2 board, with respect to our previous results. Depending on the number of processes and the load profile, we observe up to a five-fold increase in the framerate.
Buscar en Dimension