Semantic segmentation for people detection on beach images

Abstract

Cameras monitoring are increasingly aided by computer vision systems that identify risk situations. This work is part of an automatic track system to monitor beaches in the metropolitan area of Recife in order to prevent bathers to trespass the boundaries of the safe region for swimming. Semantic segmentation has gained strength in several computer vision tasks. Usually, the metaarchitecture of a semantic segmentation network consists of two modules: encoder (backbone) and decoder. This work does a study combining a set of semantic segmentation networks, Unet, Xnet, LinkNet and Unet++ with the pretrained backbones VGG16 and VGG19, to detect swimmners in beach images. We have used our own dataset, made by several images taken at the Boa Viagem beach, RecifeBrazil. The algorithms are evaluated with MIoU metric regarding the entire image scene and just in the water area. The best MIoU regarding all image was 80.87best MIoU in detecting swimmers at the beach was 85.56obtained by the LinkNet algorithm with both VGG16 and VGG19 backbones.