Real-time characteristics of marine object detection under low light conditions: Marine object detection using YOLO with near infrared camera

Abstract

This work discusses how a near infrared camera can be used to detect objects in a marine environment. The goal is to identify marine objects in real-time under low light conditions using object detection algorithm Yolo v3. Some different image processing methods were analyzed, such as saliency, edge detection and convolutional neural networks (CNN). Then the implementation of a scalable collision avoidance was presented. The system uses Intel Realsense camera, OpenDLV software framework, and the Linux operating system. For this work, an OpenDLV interface was implemented for the camera, and OpenDLV perception microservice was used to run its built-in Darknet-based Yolo v3 implementation. Then the real-time characteristics of the system and the performance were evaluated. It was proven that the system does not have real-time characteristics because of the underlying OS and sensor communication protocol. The system achieved 0.71 mean average precision (mAP) on boats with the test images. It was concluded that the system still need more complete training and testing. Finally, a suggestion on how to implement a similar system with real-time capabilities was given. This includes changing camera, OS and some part of the software that was used.