Rastreamento de múltiplos objetos utilizando modelos de aprendizado profundo em Hardware Limitado

Abstract

It is worth paying attention to the recent progress of artificial intelligence in the everyday life of people who, one way or another, use technological solutions in their activities. Much of this advancement is due to the use of convolutional neural networks, which are particularly useful in solving problems related to locating, detecting and classifying images. These networks are also used to track multiple objects. They can locate and classify an object while maintaining its unique identity over time. This is one of the reasons that make them attractive for edge computing applications, given the potential employment in areas such as electronic surveillance, traffic control, pedestrian counting, among others. On the other hand, architectures considered state-of-the-art require a lot of computing power in terms of processing, memory consumption, and thus energy. These requirements make it difficult to use complex models in hardware with limited computing resources, such as the Raspberry Pi. While it is possible to perform more complex tasks than on other platforms, using convolutional neural networks on the Raspberry Pi still presents a challenge. This monography proposes to explore the state-of-the-art YOLOv4-tiny architecture, simultaneously with the SmartSORT algorithm object tracker, in the Raspberry Pi 4. The proposed solution was tested in tracking multiple pedestrians from the MOT Challenge 2016 benchmark and in a video showing a controlled commercial environment. An average accuracy of 69% and a frame processing speed of 1.2 fps were achieved.