Research On Incremental Learning Method Of Object Detection For Unmanned Surface Vehicles

With the continuous development of the national economy,the demand for marine resources has increased,and more and more attention has been devoted to the research and development of marine technology.As a new high-tech product,the unmanned surface vehicle(USV)has great market potential and application prospects.The ability to perceive the marine environment is a key element to ensure the safety and normal operation of USV.And object detection can provide important information of sea surface object for USV to realize autonomous navigation and avoid obstacles.However,due to the diversity of sea surface objects and the complexity and changeability of the marine environment,many challenges have been brought to object detection.Therefore,the research on incremental learning of sea surface objects and incremental adaptation of marine scenes has important practical value,which can make the USV quickly adapt to the dynamic changes of the marine environment.Based on the "QZ" USV developed by Harbin Engineering University,this paper conducts research on the lightweight object detection network,incremental learning of sea surface objects,and incremental adaptation of marine scenes to improve the adaptive ability of the perception system of the USV in complex ocean environments.First of all,perception and control of the USV are usually completed on embedded devices with limited computing power.At present,most object detection algorithms have poor realtime performance on embedded devices,and cannot provide real-time and reliable sea surface object information for the USV.In order to improve the detection speed on embedded devices,this paper design a lightweight object detection network based on the YOLO v5,which can achieve a better balance between detection speed and accuracy,and is more suitable for deployment on the USV.Secondly,due to the diversity of sea surface object categories,the object detection model deployed on the USV may not be able to detect sea surface target categories that did not appear in the training dataset.While retraining the model with the new dataset will cause catastrophic forgetting,and the model cannot accurately detect the old objects.Therefore,this paper proposes an incremental learning method that can quickly learn new objects and maintain the memory of the initial objects.Based on the idea of knowledge distillation,the distillation loss function is constructed to prevent catastrophic forgetting when learning objects,so that the USV can adapt to the change of the sea surface objects and continuously learn new objects.Then,aiming at the problem that the performance of the object detection model will decrease due to changes of marine scenes such as climate and light,the domain adaptation method is used to adapt the change of marine scene.But existing domain adaptation methods aim to improve the detection performance on the target domain,ignoring the sharp drop of the model’s performance on the source domain.In order to solve the problem of catastrophic forgetting when learning new domains,an incremental domain adaptation method is proposed based on the above incremental learning method,which can be divided into two stages: in the first stage,the source domain images are transformed into images that match the data distribution of the target domain;in the second stage,the above incremental learning method is used to learn the generated dataset.It enables the model to learn new marine scenarios while avoiding significant performance degradation in the learned scenarios,and improves the adaptive ability of the USV perception system.Finally,based on the "QZ" USV platform,experiments were carried out in the real marine environment.And the proposed methods were verified on real marine images that captured by the photoelectric pod carried by the USV.The experimental results show that the lightweight object detection model meets the real-time requirements of the USV.The proposed incremental learning method can continuously learn new sea surface objects and avoid catastrophic forgetting of initial objects;The proposed incremental adaptation method can adapt to the dynamic changes of the marine scenes.These methods are suitable for deployment in the perception system of USVs,providing more reliable sea surface object information for the autonomous driving,and improving the intelligence level of the USV.