Real-Time Object Detection for Unmanned Underwater Vehicles Using Movidius Neural Compute Stick

Deep learning-based object detection methods have demonstrated promising results in controlled environments. Due to specific limitations, these methods are insufficient for unstructured environments, such as underwater exploration or espionage using unmanned underwater vehicles (UUV). Poor image quality is the primary issue limiting UUV’s functionality. Underwater objects are typically small and blurry as a result of severe noise that confuses the detectors. This study presents a graph convolution neural network (GCNN)-based real-time object tracking system in an unstructured environment. On the algorithm side, object’s regions are categorized based on the pixel population density. Resultantly, the regions are disintegrated recursively into an image patch tree. Moreover, the surfaces are subsequently and proficiently stored, later retrieved as patches. To keep the power consumption low, a PYNQ Z2 FPGA is used as development platform in connection with a Movidius NCS stick for the outsourcing of the GCNN. The proposed technique is generic and reducing high computational complexity and power consumption.