Embarcações autônomas na região amazônica: aplicação de técnicas de aprendizado profundo para detecção de objetos localizados na superfície de rios

Abstract

The application of Computer Vision and Deep Learning techniques in water surface object detection context has been emerged as a strong trend in autonomous vessels scena rio. In this work, we present performance comparisons between different object detection models by using two distinct datasets: WSODD (Water Surface Object Detection Data set) and WSOD-ARD (Water Surface Object Detection - Amazon Rivers Dataset). The first one, WSODD, is characterized by being publicly available, wide and contains ob jects located on marine water surfaces. The second one, WSOD-ARD, was created and annotated with objects belonging to Amazon rivers water surface scenario. For each one of these datasets, we used YOLOv5 algorithm as architecture for training water surface object detection models. The first model, trained by using WSODD dataset, reached a mAP of 76.3 %, outstanding in 11.3 % the mAP obtained by CRB-Net detector in this benchmark dataset. The second model, trained by using WSOD-ARD dataset, reached a mAP of 75.4 % by using transfer learning techniques. Finally, this model was deployed into an edge device (Nvidia Jetson Nano embedded plataform) in order to simulate a real scenario the proposed application.