Real-time pose estimation for an underwater object combined with deep learning and prior information

Abstract

At present, the underwater autonomous operation based on monocular vision has poor accuracy and low intelligence, due mainly to the low accuracy of pose estimation. To solve this problem, we propose a real-time pose estimation method for underwater cylinders and cuboids. The first challenge in processing underwater images is image degradation, which is overcome by using a scale-optimized dark channel prior dehazing algorithm. The lightweight improved You Only Look Once v5 is used to obtain the pixel information of the four control points and obtain the bounding box close to the edge of the object, which makes the pose estimation more accurate. We then propose an underwater optical imaging model that overcomes the challenges posed by refraction. Finally, the improved algorithm based on the perspective-n-point problem is used to estimate the pose of the object in real time. We deployed the algorithm in the edge computing device NVIDIA Jetson TX2 and achieved excellent performance. The experimental results show that our method can achieve high-precision monocular pose estimation without producing a large-scale pose dataset, and can be used to provide reliable information for underwater autonomous operation tasks.