【作者】 王浩；

【导师】 范菁； 董天阳；

【作者基本信息】 浙江工业大学 ， 计算机科学与技术， 2016， 硕士

【摘要】 随着计算机图像处理技术的发展,基于视频的车辆检测技术在智能交通管理系统的研究中发挥了重要的作用。白天的车辆检测技术已经相当完善,得到了广泛的应用,但长期以来由于夜间视频的质量不高,基于视频的夜间车辆检测技术研究相对受到了很多制约。近年来监控设备不断普及,图像技术快速发展,交通视频的拍摄效果越来越清晰,夜间视频的质量也得到了很大的提升。但是由于夜间场景复杂,可参照物少,光线太弱等原因,适用于白天车辆检测的算法在夜间并不适用。所以为了有效检测和识别夜间高速公路中的车辆,本文针对夜间车辆的运动信息和夜间车辆的局部特征信息进行了相关的研究,主要的研究内容如下:(1)设计了改进的夜间车前灯提取算法。由于直接提取视频每一帧中的车灯会存在很多噪点,所以在车灯提取之前首先进行中值滤波对图像做平滑操作,过滤掉视频中一部分噪点。然后统计夜间场景感兴趣部分(ROI)亮度分量直方图,结合得到的车灯亮度先验知识,利用一维最大熵阈值法获取自适应的阈值,对图像进行分割提取得到车灯。(2)提出了基于空间特性的车灯分组方法。考虑到同一辆车的车灯在面积、宽高比、质心位置等特征存在相似之处,本文结合车灯的空间特性和车辆的特征来定义规则完成车灯匹配,该方法不仅对于一辆车有多于两个车灯的情况有很好的匹配效果,而且对于并排同速同向的车辆也有很好的处理效果。(3)实现了基于Kalman滤波器的车辆跟踪算法。针对夜间高速视频中邻近帧间车辆变化较为稳定的情况,本文利用车灯对的质心和车灯包围盒的信息作为车辆匹配的特征。采用Kalman滤波器对其建立模型,根据当前帧中车灯的特征信息,对下一帧中车辆可能出现的位置进行预测,使用检测到的车辆信息更新模型,不断迭代该过程,完成车辆的跟踪。(4)提出了基于BOF的夜间高速车辆类型识别算法。本文提出了一种改进的基于视觉词袋的车型分类算法,首先使用SURF对夜间车辆的特征进行描述;然后使用基于最大最小距离算法改进的K-means聚类方法对特征点进行聚类,形成视觉词典;最后使用SVM来完成分类。实验结果可以看出该方法对夜间车辆的识别有较高的准确率和较快的速度。更多还原

【Abstract】 With the development of computer image processing technology,video based vehicle detection technology plays an important role in the research of intelligent traffic management system.Daytime vehicle detection technology is already quite perfect,and has been widely used,but the nighttime video has low quality for a long time,so video based vehicle detection technology in the night has a lot of constraints.In recent years,with the rapid development of computer image processing technology and monitoring devices,traffic videos are becoming more and more clear,and night video quality has also been greatly improved.Due to the complexity of the night scene,the reference substance is too few and the light is too weak,so the vehicle detection algorithm that is suitable for daytime cannot be applied to nighttime.According to the above-mentioned problems this paper researched on the motion information of vehicles and local feature information of vehicle at night.The main research results are as follows :(1)An improved algorithm for the extraction of the vehicle headlights was designed.There will be a lot of noise when we directly extract the light on video,so before the lights are extracted,the median filtering is done to smooth the image,so that filtering out the noise in the video.Then we can get a priori knowledge of the brightness of the headlights based on statistical knowledge,then using one dimensional maximum entropy threshold method to obtain the adaptive threshold.The threshold is used to segment the image,and then extracting the bright block as a candidate lamp,finally filtering out the unstable bright blocks to get the car lights.(2)A light grouping method based on spatial characteristics is proposed.Consider that the lights’ characteristics of the same car are very similar,such as the area,aspect ratio,the location of the center of mass and so on.This paper combines the spatial characteristics of the lights and vehicle characteristics to define rules for lights matching.This method not only has a very good matching effect for a car with more than two car lights,but also has a good effect on the same direction of the vehicle with the same speed.(3)A vehicle tracking algorithm based on Kalman filter is realized.Because the change of the vehicle is stable between the adjacent frames in the night high speed video,in this paper,we use the information of the center of mass and the surrounding of the vehicle to match the vehicle.Kalman filter is used in the design of the model,and according to the characteristic information of the vehicle lamp in the curre nt frame,the possible position of the vehicle in the next frame is predicted.Using the detected vehicle information to update the model,and then repeat ing the process to complete the vehicle tracking(4)An algorithm based on BOF is proposed that identifies the type of high speed vehicle at night is proposed.This paper presents an improved algorithm for vehicle classification based on visual word bag.At first using SURF to describe the characteristics of the vehicle at night;then the feature points are clustered by the improved K-means clustering algorithm based on the max min distance algorithm to form a visual dictionary,finally using SVM to complete the classification.The experimental results show that the method can identify the vehicle at night with high accuracy,and the processing speed is very fast.更多还原