【作者】 王洋；

【导师】 朱力强；

【作者基本信息】 北京交通大学 ， 载运工具运用工程， 2019， 博士

【摘要】 基于视频智能分析的异物入侵识别技术已经成为高速铁路周界安全检测领域的重要发展方向,目前制约视频分析技术广泛应用的难题之一是报警区域的自动辨识问题。为了实现高速铁路场景报警区域快速、准确、自动地分割与识别,论文提出了基于多特征通道、自适应检测算子的图像快速分割算法和基于极简化卷积神经网络的局部区域快速识别算法,在计算速度、分割精准度、识别准确率等多个方面达到了有效的平衡。论文首先针对现有区域分割算法计算量过大的问题,通过分析铁路场景直线特征强烈的特点,提出基于多特征融合与优化梯度检测算子的区域快速分割算法。该算法首先利用霍夫变换寻找铁路场景直线特征极大值来调节用于提取图像特征梯度分布的检测算子;然后以少量最优的自适应检测算子来同时计算像素点色彩纹理特征分布以及像素点相似度分布,获取精准的、带有权重的边界点,并进一步通过自动筛选强弱边界点来防止图像分割结果过度碎片化;最后设计基于边界点权重及碎片区域面积的组合规则来将碎片区域快速组合成局部区域。对比实验结果表明,所提算法能够快速、有效的分割图像得到局部区域精确边界,为后续的区域识别部分奠定良好的基础。针对现有区域识别卷积神经网络计算量大、占用内存多、需要GPU实时处理的问题,论文提出了基于预训练的卷积核权值优化算法。该算法通过构建浅层自编码神经网络,自动学习铁路不同区域特征模板,解决了铁路场景同类物体差异性较大、同一物体图像成像质量受光照变化及相机因素影响较大的优化建模问题。对比实验结果表明,所提算法有效提高了传统卷积神经网络对铁路场景在局部区域的识别准确率,也为类似应用问题提供了解决思路。为进一步加快铁路场景分割与识别的计算速度,论文提出一种两段式区域分割算法。在第一阶段,利用极简化卷积神经网络快速扫描铁路场景并确定轨道区域的潜在位置;在第二阶段,对潜在位置进行精确分割与识别。针对极简化的卷积网络准确率较低的难题,提出一种卷积通道稀疏化的代价函数来优化训练过程,提高特征图的差异性。对比实验结果表明,两段式算法极大减少了全景图像中无关区域的冗余计算时间,极简化的卷积网络的准确率也得到了有效弥补,有效地提高了系统的工作效率和性能。与现有技术相比,论文提出的系列算法对铁路场景分割与识别的像素级准确率大大提高,计算时间更短,网络参数更少,摆脱了对GPU显卡的依赖,降低了应用系统成本,便于移植到不同配置的数据处理平台,目前已经在沪宁城际高速铁路上得到实际应用。更多还原

【Abstract】 The intrusion object detection technology based on video intelligent analysis has become a vital development direction in the field of high-speed railway perimeter security detection,one of the problems that currently restricts the wide application of video analysis technology is the automatic identification of alarm areas.To achieve a fast,accurate,and automatic segmentation and recognition,a fast image segmentation algorithm based on adaptive feature gradient detector and a fast image recognition algorithm for the local area based on simplified convolutional neural network are proposed,and an effective balance has been reached in the areas of calculation speed,segmentation precision,recognition accuracy,parameter calculation amount and manual participation.Firstly,aiming at the large calculation caused by the existing segmentation algorithm,the strong linear characteristics of the railway scene is analyzed and a fast segmentation algorithm based on multi-feature fusion and the optimization of adaptive feature gradient detector is proposed.This algorithm first uses the Hough transform to identify the maximum value of the linear feature of the railway scene to adjust the feature gradient detector for extracting the gradient distribution of the image.Then,a small amount of optimized detectors are applied to simultaneously calculate the pixel color&texture feature distribution,and the pixel similarity distribution to obtain an accurate boundary point with weight,further preventing fragmented regions in image from being excessively fragmented by automatically screening strong and weak boundary points;finally,a combination rule is designed to quickly combine the fragmented regions into local areas based on the boundary point weight and the size of the fragmented regions.The comparison experiment results show that the proposed algorithm based on feature-fusion and adaptive gradient detector can extract the precise boundary of local areas quickly and effectively,which lays an optimal foundation for the subsequent region recognition.Aiming at the problem that the existing local area identification by convolutional neural network is computationally intensive,and occupies a large amount of memory,and requires real-time processing by the GPU,a pre-training-based convolution kernel weight optimization algorithm is proposed in this paper.By constructing a shallow autoencoder network,the algorithm learns the basic feature templates of different regions of the railway automatically,the pre-trained kernels will solve the model optimizing problem caused by the various feature in the same category and the bad image quality of the same object when the illumination and camera factors changed.The comparative experimental results show that the proposed algorithm has effectively improves the recognition accuracy of the traditional convolutional neural network in the partial region of the railway scene,and it also provides solutions for similar application problems.To further accelerate the computational speed of railway scene segmentation and recognition,this paper proposes a two-step region segmentation algorithm.In the first phase,the ultra-lightweight convolutional neural network is used to quickly scan the railway scene and determine the potential location of the track area.In the second phase,the potential location is accurately segmented and identified.Targeting at the problem of low accuracy of ultra-light quantitative convolutional network,a sparse cost function of convolution channel is proposed to optimize the training process and enhance the difference of feature map.The comparative experimental results show that the two-stage algorithm greatly reduces the redundant computing time of the panoramic image,and the accuracy rate of the ultra-lightweight convolution network is effectively compensated,which greatly improves the system’s working efficiency and performance.Compared with the existing technology,the series of algorithms proposed in this paper greatly improve the pixel-level accuracy of railway scene segmentation and recognition,the calculation time is comparatively shorter,the network parameters are less,the reliance on GPU graphics card is eliminated,the application system cost is reduced,and easy to migrate to different configurations of data processing platform.Currently,this algorithm is being put to practical use on the Shanghai-Nanjing Intercity High-speed Railway更多还原