面向无人艇智能感知的水上目标识别算法研究

针对水面无人艇(unmanned surface vessel,USV)智能感知系统对图像处理过程的准确性和实时性要求,研究了一种根据无人艇上机载视觉传感器对水上目标进行识别与定位的算法.首先根据开源数据集与实验数据采集图像,对实验数据抽帧、去重、标注、统计,创建了水上目标识别数据库YZ10K;其次实践了主流的基于深度学习的目标检测方法,包括Faster R-CNN、SSD、YOLOv3等;最后针对水上目标特点,提出了一种基于改进YOLOv3的增强型轻量级水上目标检测网络WT-YOLO(water target-you only look once).无人船实验验证表明,WT-YOLO算法取得了准确且快速的目标识别效果,平均精度为79.30％,处理速度为30.01 fps.

Research on water target recognition algorithm for unmanned surface vessel

In this paper, a water-target recognition algorithm based on the data acquired by the onboard visual sensor from unmanned surface vessels (USV) is reported, in order to satisfy the accuracy and speed requirements of USV intelligent sensing system. The main outcome are summarized as follows: First, images are collected based on open source datasets and experimental data, to create a watertarget recognition database which named YZ10K; second, popular deep-learning based target detection methods including Faster RCNN, SSD, YOLOv3, etc. are implemented and compared; third, based on the characteristics of water targets, an enhanced lightweight Water Target detection network WT-YOLO (water target-YOLO) is proposed. The experimental verification shows that the WT-YOLO algorithm based on improved YOLOv3 has achieved accurate and real-time target recognition with the mean average precision (mAP) of 79. 30% and frame per second of 30. 01.