【摘要】 为解决背景建模等传统视频目标识别算法在内河水运复杂环境误差过大的问题,提出了甚高速区域卷积神经网络的船舶识别检测方法.分析了传统方法不足,阐述了卷积神经网络及后续的区域卷积神经网络的机制,给出了甚高速区域卷积神经网络特征模型,解析了损失函数的参数构建、参数设定,设定候选区域网络预测目标边界、计算匹配目标概率.经实际内河运动船舶视频检测表明,该算法对船舶识别率优于90%,同时对不同清晰度、不同视角、不同船舶流量的场景具有很好的鲁棒性,比传统的背景建模算法提高25.75%.更多还原

【Abstract】 To avoid the excessive error during the implementation of background modeling and other traditional video object recognition algorithms in complex inland waterway environment,a new ship identification detection method based on faster region convolutional neural networks( i. e. Faster R-CNN) was proposed. The shortcomings of these traditional methods were analyzed,elaborating the mechanism of convolutional neural network and the subsequent regional convolutional neural network was also described. A model of R-CNN was put forward,which worked out how to construct and set the parameter of Loss Function,set region proposal networks( RPN) to predict a target boundary,and calculated the probability of matching targets. The actual video detection for moving inland river ships indicated that the ship identification algorithm held a ship recognition rate of over 90%. Meanwhile,this new algorithm had good robustness,25. 75% higher than traditional background modeling algorithm in situations with different visual clarity,from different perspectives,and regardless of the number of ships.更多还原