Underwater Drowning People Detection Based On Bottleneck Transformer And Feature Pyramid Network

Abstract

Emergency rescue is a critical security guarantee for transportation and engineering operations at sea. As an important equipment for the emergency rescue mission, under-water robots could detect drowning people automatically, which can effectively improve operational efficiency. Due to the severe attenuation of light underwater, optical imaging devices such as cameras cannot obtain information from long-distance scenarios. This paper proposed an object detection network based on forward looking sonar images. Due to the performance limitations of the processer in the robots, the proposed network is designed to be lightweight based on the Bottleneck Transformer and Feature Pyramid Networks to detect drowning people underwater instead of large networks. The experimental results show that the proposed network can ensure relatively good accuracy with fast detection speed.