改进深度学习算法PSPnet在水上障碍物检测中的研究

Research on application of improved deep learning algorithm PSPnet in water obstacle detection

无人艇是可用于港口、码头巡逻、海上环境检测的自主机器人,水上障碍物检测作为无人艇进行环境感知的关键能力,对无人艇有重要影响。针对无人艇水上障碍物检测,文中提出一种基于改进PSPnet框架的水上障碍物检测算法,采用Mobilenetv2中的轻量化结构Bottleneck作为主干特征提取网络,提高实时性,同时分别采用加强特征提取,提高分辨力的策略提高精度。采用海洋数据集Master1325作为训练集,MODs作为评价集,结果显示,在MODs上的精度F-measure达到86.8,在MODs中危险区域的F-measure为81.1,且在NVIDIA GTX 2080Ti上的图像推理速度达到61 FPS,检测速度和精度较高,证明了所提出算法的有效性。

Unmanned surface vehicle is an autonomous robot,which can be used in port and dock patrol and marine environment detection.As the key ability of unmanned surface vehicle to sense the environment,obstacle detection on water is important to unmanned surface vehicle.As for the detection of obstacles on the water of unmanned surface vehicles,an algorithm for detecting obstacles on the water based on the improved PSPnet framework was proposed.The lightweight structure Bottleneck in Mobilenetv2 was used as the main feature extraction network to improve the real-time performance.Meanwhile,the strategies of strengthening feature extraction and improving resolution were adopted respectively to improve the accuracy.Taking the marine data set Master1325 as the training set and MODs as the evaluation set,the results show that the accuracy of F-measure on MODs was 86.8,that of dangerous area in MODs was 81.1,and that the image reasoning speed on NVIDIA GTX 2080Ti was 61 FPS,demonstrating that the proposed algorithm is effective.