论文部分内容阅读
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%.
In order to solve the problem of traditional video target recognition algorithm such as background modeling and so on, which has a large error in inland waterway transport complex environment, a method of ship identification and detection of very high speed regional convolution neural network is proposed.The traditional method is analyzed and the convolution neural network and Followed by regional convolution neural network mechanism, given very high speed regional convolution neural network characteristic model, the analysis of the loss function parameter construction, parameter setting, set the candidate regional network prediction target boundary, calculate the probability of matching target. The actual video detection of inland waterway ships shows that this algorithm is better than 90% for ship recognition rate and has good robustness to scenes of different sharpness, different perspectives and different ship traffic, which is 25.75 higher than that of traditional background modeling algorithm %.