Real-Time Underwater Onboard Vision Sensing System for Robotic Gripping

In this study, a real-time underwater onboard vision sensing system is developed for robotic gripping. First, an efficient image enhancement method based on the Retinex theory is presented. The enhanced images are provided for underwater robot to observe the seabed environment clearly via cameras. Subsequently, a real-time lightweight object detector (RLOD) for the mobile embedded platform is proposed. The RLOD is designed as an hourglass detector network, which introduces dense connections and a featured pyramid network to improve the detection performance and speed. Moreover, from an engineering perspective, two merging methods are used to deploy the trained network. It can be implemented at 11.11 frames per second (FPS) on the Nvidia Jetson TX2 processor, satisfying the real-time requirement of underwater robotic gripping. Furthermore, a refraction tracing model is constructed. The comparative results show the effectiveness of the proposed methods. Finally, this onboard vision sensing system is mounted on an underwater robot with a manipulator to implement robotic gripping. Pool and sea experiments are conducted to verify the practicability of the developed system.