Abstract:Object detection based on visual images provides significant technical support to realize autonomous environment perception of unmanned surface vehicles (USVs), yet difficulties exist for object detection on the sea surface, such as the constantly-changing weather conditions, considerable changes in the scale of objects, and severe object shaking. In this paper, an USV object detection algorithm is proposed based on an optimized feature fusion network. YOLOv4 was selected as the baseline, which is currently one of the object detection models with the best trade-offs between speed and accuracy. In the proposed weighted Cross Stage Partial Path Aggregation Network (wCSPPAN) structure, the concept of the Cross Stage Partial Network (CSPNet) was applied to the feature fusion network, so as to significantly reduce the amount of computation. In addition, weights that can be learned were introduced to learn input features of different resolutions in a more reasonable way. Finally, to obtain cross-space and cross-scale feature interactions, attempts were made to integrate the universal visual component of the Feature Pyramid Transformer (FPT) with the transformer mechanism in the YOLOv4 model. The FPT was used to transform the feature pyramid into another feature pyramid of the same size with richer contextual information for each level of feature map. We compare the performance with other advanced object detection algorithms. The experiments conducted on the dataset of sea surface buoys, which was collected and produced independently for the present study, and the open source Singapore Maritime Dataset verify that the proposed method achieved good results in the detection of sea objects of different scales under various weather conditions. In particular, the detection of long-distance small objects under foggy conditions with low visibility and extreme conditions was improved, and real-time sea surface object detection was achieved with an inference speed of up to 36FPS on a single RTX 2080Ti. Further, the detection results on the COCO dataset and KITTI dataset verify the excellent generalization ability of our proposed object detection model.

Fast Object Detection from Unmanned Surface Vehicles Via Objectness and Saliency

Object Detection Based on Saliency and Sea-Sky Line for USV Vision

A Method with Improved Accuracy and Robustness for Object Detection in Wharf Scenarios

Visual Detection Algorithm for Enhanced Environmental Perception of Unmanned Surface Vehicles in Complex Marine Environments

Few-Shot Object Detection Using Multimodal Sensor Systems of Unmanned Surface Vehicles

A Real-time Algorithm for Visual Detection of High-speed Unmanned Surface Vehicle Based on Deep Learning

A 3D Object Detection Based on Multi-Modality Sensors of USV

Research on Unmanned Vessel Surface Object Detection Based on Fusion of SSD and Faster-RCNN

Real-Time Obstacle Detection for Unmanned Surface Vehicle Maneuver

Improved Object Detection Method for Unmanned Surface Vehicle Using Real-Time Neural Networks

Object Detection And Tracking Of Unmanned Surface Vehicles Based On Spatial-Temporal Information Fusion

Unmanned Surface Vessel Visual Object Detection under All-Weather Conditions with Optimized Feature Fusion Network in YOLOv4

A object detection and tracking method for security in intelligence of unmanned surface vehicles

Global sparsity potentials for obstacle detection from Unmanned Surface Vehicles

A Fusion Algorithm of Object Detection and Tracking for Unmanned Surface Vehicles

Research on fast segmentation algorithm of feasible region and obstacles of unmanned surface vessels

Lightweight object detection algorithm based on YOLOv5 for unmanned surface vehicles

YOLO-MRS: an Efficient Deep Learning-Based Maritime Object Detection Method for Unmanned Surface Vehicles

Vision based long range object detection and tracking for unmanned surface vehicle

Obstacle Detection of Unmanned Surface Vehicle Based on Lidar Point Cloud Data

A Multi-modality Sensor System for Unmanned Surface Vehicle