Marine Litter Classification Through Multi-Camera System: Perception and Latency

This work presents a pilot study to classify marine litter observed through a distributed static camera system. The acquisition from the multi-camera sources located at different viewpoints transmits data to the edge devices for independent, low-latency image classification. We employed transfer learning-based networks to monitor moving objects using a dataset sourced from the JAMSTEC library of Deep-Sea images for training purposes. The pretrained VGG19, ResNet50, and DenseNet121 network architectures are employed as fixed feature extractors to train on underwater marine dataset. The achieved accuracies on the JAMSTEC test data are 78%, 83%, and 84% for VGG19, ResNet50, and DenseNet121, respectively. We further simulated, through the experimental setup, the impact of latency on the performance of classification output when the video streams from two cameras arrive at the edge device with different delays.