A Novel Garbage Detection Algorithm for Unmanned Surface Vehicles based on Reparameterized Model and Dual Path Feature Fusion

Abstract

With the growing garbage flowing into the ocean through inland rivers, garbage detection and cleanup has become an urgent and necessary task for the safety of the entire ecosystem. Considering the expensive labor cost, Unmanned Surface Vehicle(USV) has been accepted as an important intelligent robot in river management field, and camera is widely adopted as a cost-effective way compared to other sensors like radar and laser for USVs in garbage detection. However, garbage detection based on vision is often affected by factors, such as the small size of distant targets, surface glare, and interference from other floating objects. It is a conflicting issue to achieve a higher detection accuracy with limited resources. To solve these problems, a novel detection algorithm is put forward for USV in garbage cleanup in the paper. Specifically, Dual Block Module(DBM) and QarepC3 modules are proposed based on the reparameterized model and dual-path feature fusion approach. EMIOU and Joint Attention Module(JAM) are constructed according to the characteristics of floating garbage. Subsequent to comprehensive evaluations, the proposed network exhibits not only high detection accuracy and computational efficiency but also robust performance in the complex environments of inland river. Moreover, it outperforms state-of-the-art networks in surface garbage detection for USV in the experiments.