基于深度学习的多船舶目标跟踪与流量统计
2020-04-22冼允廷邱伟健
冼允廷 邱伟健



摘 要:研究船舶的目标跟踪对提高水上目标视频图像智能监管的水平有着至关重要的作用,系统通过深度学习SSD模型进行对船舶目标定位检测,使用修正的KCF算法对检测到的船舶目标进行跟踪。把深度学习的方式引入船舶目标检测领域,与传统检测方法相比精准率大大提高,同时提出了一个修正的KCF算法对多船舶目标进行跟踪,较好地解决了目标漏检与重复统计的问题。对大量船舶目标样本进行训练学习后,船舶检测定位精准,检测成功率达到91%以上,船舶跟踪算法快速稳定,检测与跟踪算法达到30帧每秒,船舶目标流量统计准确率达到95%以上,整个系统框架满足实时性的要求。
关键词:船舶跟踪; 船舶检测与统计; 深度学习; 水上交通
中图分类号: U664.121
文献标志码: A
Abstract:The research on ship tracking from video surveillance plays a vital role in improving the intelligence of maritime safety regulation. We applied a deep learning method-SSD to detect and locate ship targets, then proposed a modified KCF algorithm to track and count these targets. By introducing deep learning methods, the accuracy of ship detection is greatly improved comparing with the traditional methods. An improved KCF algorithm can realize to track more than one ships. It solves the problem of ship detection and statistics and without missing. Benefits from the modified KCF algorithm, omission and duplication failure are effectively relieved while maintaining real-time processing speed. According to the experiments we performed, the accuracy of traffic flow data reaches 95%.
Key words:Ship tracking; Ship detection and statistics; Deep learning; Water traffic
0 引言
船舶目标跟踪与流量统计是水上交通智能化研究的一个重点,一直以来都是以雷达及AIS系统作为水上船舶目标管理信息化的主要手段;近年来由于水上高清视频监控发展的越来越快,视频图像更加让监管人员能直观地了解现场情况,有利于现场事故救援,方便违规事件调查取证等,所以基于视频图像的智能监管系统也越来越受到水上监管部门的重视。目前在国内外对船舶航道方面的智能监控平台并不多见,但由于水上交通运输量大,运费较陆上便宜,水上交通也日渐繁忙,因此建立一套能够对船舶检测、跟踪、统计及识别等智能化的视频监控平台,对减少交通事故,保障航行安全,提高信息化和智能化水平有着深遠的意义和广阔的前景。……
