Towards Real-Time Human Detection in Maritime Environment Using Embedded Deep Learning

Marine search and rescue missions necessitate a lot of effort and expenses. The use of technological advancements facilitates discovering and locating individuals and aids in the directing of rescuers and medical teams. This has the potential to save human lives while also lowering costs. The characteristics of the marine environment create additional challenges for computer vision techniques used to detect the presence of human in a scene. Currently, artificial intelligence (AI) techniques based on convolution neural networks (CNNs) provide solid solutions to detect and locate objects. In this paper, the relevance of the emergent You Only Look Once (YOLO) in detecting humans in maritime environment is investigated. The available models of YOLOv4 are trained using a custom dataset. The trained models are evaluated using recognized evaluation parameters. In addition, the inference speed is reported targeting embedded low-power hardware platforms dedicated for AI applications.