Research on Real-Time Diver Detection and Tracking Method Based on YOLOv5 and DeepSORT

In this paper, we explore a class of diver detection and tracking method to assist the diver to collaborate with underwater robot in the deep sea. At first, we employed YOLOv5 to train a diver detection model on the VDD-C dataset. And then we compare the detection accuracy and speed between model with different backbones. We transplant the trained model into an embedded platform, and discover the detection speed is too slow to meet the need of real time detection. Therefore, we replaced the original backbone network with MobileNetV3 which is a light network designed for embedded device, and the detection speed is twice as faster as before on the Jetson nano. In the last, we combine YOLOv5 with DeepSORT tracking algorithm to track diver. The tracker will assign an id above the bounding box, which can empower robots to distinguish each diver. At the same time, we analyzed a specific scene that occurs in the deep sea: a diver is swimming forward in front of robot, and bubbles due to breathing obscure most of diver's body to make detector lose target. And when the bubbles disappear, the tracker track the diver again and assign the same id with before.