智能交通系统中行人检测算法的研究

发布时间：2018-02-21 02:22

本文关键词： 智能交通行人检测特征融合卷积神经网络　出处：《哈尔滨理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着科技水平的不断发展,智能交通技术愈加完善,其应用场景及需求量不断扩大。相较于传统的智能交通系统,新一代智能交通系统无论在硬件基础上还是在软件算法上都日趋完善,已经具备实时性、及时性、高准确性、低误报率等优质性能。新一代智能交通系统通过对路况进行全天候、无间断监测,同时对实时获取的监测数据通过计算机技术进行高效计算处理,以此作为依据实现道路交通智能化。行人检测模块作为智能交通系统中重要的组成部分已经成为国内外学者研究的核心课题,在科研领域及工程应用方面均具有广泛的发展前景。本文针对智能交通系统中的行人检测部分深入研究,通过对高清摄像机拍摄的行人图像进行实验分析。与当前流行的行人检测算法进行实验对比进而深入分析本文提出的改进算法。在行人检测领域中,检测算法方面的工作内容主要集中在以下三个方面:一,通过图像滤波、形态学处理等相关图像处理技术,对视频图像序列进行优化处理,在保留图像有效信息的基础上尽量减少冗余信息的干扰;二,特征集的选取,本文区别于传统特征集的选取,将单一的梯度方向直方图(Histogram of Oriented Gradient,HOG)特征与局部二值(Local Binary Patterns,LBP)特征利用统计直方图级联进行特征融合,以获取更深刻更全面的图像信息;三,在训练分类器的部分,改变传统的训练方式即利用线性支持向量机(Support Vector Machine,SVM)训练分类器,采用截取后的卷积神经网络(Convolutional Neural Network,CNN)作为分类器的训练算法,利用三层全连接层进行训练,最终获取性能较好的分类器。至此,通过对以上三方面的研究与改进,优化了算法的实时性与鲁棒性。同时也降低了行人背景复杂时检测失败的出现几率,提高了行人检测的有效性与可靠性。
[Abstract]:With the continuous development of science and technology, intelligent transportation technology is becoming more and more perfect, and its application scene and demand are expanding. Compared with the traditional intelligent transportation system, The new generation of intelligent transportation system is becoming more and more perfect on the basis of hardware and software algorithm. It has the characteristics of real-time, timeliness and accuracy. The new generation of intelligent transportation systems can monitor the traffic conditions all the time, without interruption. At the same time, the monitoring data obtained in real time can be efficiently calculated and processed by computer technology. As an important part of intelligent transportation system, pedestrian detection module has become the core research topic of domestic and foreign scholars. In the field of scientific research and engineering applications, there is a wide range of development prospects. In this paper, the intelligent transportation system in the pedestrian detection part of in-depth research, Through the experimental analysis of the pedestrian images taken by the high-definition camera, and compared with the current popular pedestrian detection algorithms, the improved algorithm proposed in this paper is analyzed in depth. In the field of pedestrian detection, The work of the detection algorithm is mainly focused on the following three aspects: first, through image filtering, morphological processing and other related image processing technology, the video image sequence is optimized. On the basis of preserving the effective information of the image, the interference of redundant information is reduced as far as possible. Secondly, the selection of feature sets is different from the traditional feature set selection. The histogram of Oriented gradient histogram is fused with the local binary local Binary patterns to obtain more profound and comprehensive image information. Third, in the part of the training classifier, To change the traditional training method, we use linear support Vector machine (SVM) to train classifier, adopt convolutional Neural network (CNN) after intercepting as the training algorithm of classifier, and use three layers full join layer to train. Finally, the classifier with better performance is obtained. By studying and improving the above three aspects, the real-time and robustness of the algorithm is optimized. At the same time, the probability of detection failure is reduced when the pedestrian background is complex. The effectiveness and reliability of pedestrian detection are improved.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;U495

【相似文献】