基于压缩域车辆异常事件检测

发布时间：2018-06-18 01:27

本文选题：HEVC + 运动矢量　；参考：《江南大学》2017年硕士论文

【摘要】：目前对于交通事故的采样和调查主要通过人工查看的方式,这种方法费时费力。近年来计算机视觉技术的发展为车辆异常事件的检测提供了诸多解决方案,但都是基于像素域的方法。然而由于监控视频的高分辨和像素域方法的解码特性,使得传统像素域方法很难满足实时性的要求。在实时性上压缩域有着很大优势。通过对当前最新压缩域编码标准HEVC的研究和大量交通事故视频的分析,发现HEVC包含的分块模式和运动矢量可以看作是目标边界和光流的粗糙分析,这些信息对于交通事件的检测很有帮助。当前在压缩域方向的研究主要是对运动目标的分割和跟踪,这些压缩域算法追求高分割精度而使得计算复杂度较高,且国内外对于交通事件很少有研究。本文在研究HEVC编码标准的基础上,根据HEVC的编码特性,提出一种根据运动矢量和块划分信息的运动强度值算法实现对运动目标的检测,并在此基础上提出一种多参数模糊逻辑算法检测出车辆异常事件。本文的主要工作和创新点包括:(1)针对HEVC的编码结构和视频内容有很大相关性的特点,提出一种基于运动强度值的运动目标检测方法。首先从HEVC压缩码流中提取出运动矢量和预测单元块划分信息,对运动矢量进行去噪和归一化的预处理,使得运动矢量更加平稳可靠,然后结合块划分信息计算出每一个最大编码单元块的运动强度值。最后通过设定的自适应阈值将运动强度值较小的单元块滤除,可以得到运动剧烈的运动目标,实现本文算法中基于运动强度值的运动目标检测。(2)当前像素域方向关于交通事故检测算法主要依据质心坐标距离差和时间差进行判断,针对这些算法计算复杂度高且可靠性较差,本文通过对车辆所在区域面积和运动强度值变化的分析,提出一种多参数模糊逻辑理论来检测车辆异常事件。首先利用广度优先搜索算法对含有运动强度值的最大编码单元块构建连通区域,提取出连通区域的面积和其对应的运动强度值。其次根据面积梯度信息区分出不同车型和交通冲突场景。最后通过每一帧的运动强度值、面积梯度等级和运动强度离均差值构成模糊逻辑理论的输入参数,根据车辆运动特征隶属度函数判断当前帧是否发生车辆异常事件。通过对25组交通事故视频的实验验证,可以精准地检测出其中22组交通事故,检测率为88%,证明了本文算法的有效性和可靠性。
[Abstract]:At present, the sampling and investigation of traffic accidents are mainly done by manual inspection, which is time-consuming and laborious. In recent years, the development of computer vision technology has provided many solutions for vehicle anomaly detection, but all are based on pixel domain methods. However, because of the high resolution of surveillance video and the decoding characteristics of pixel domain method, the traditional pixel domain method is difficult to meet the real-time requirements. In real-time compression domain has a great advantage. Based on the research of the latest compression domain coding standard HEVC and the analysis of a large number of traffic accident videos, it is found that the block pattern and motion vector contained in HEVC can be regarded as the rough analysis of the target boundary and optical flow. This information is helpful for traffic incident detection. At present, the research on the direction of compressed domain is mainly on the segmentation and tracking of moving objects. These compressed domain algorithms pursue high segmentation accuracy and make the computation complexity higher, and there are few researches on traffic events at home and abroad. In this paper, based on the research of HEVC coding standard and according to the coding characteristics of HEVC, a motion intensity algorithm based on motion vector and block partition information is proposed to detect moving targets. On this basis, a multi-parameter fuzzy logic algorithm is proposed to detect the abnormal events of vehicles. The main work and innovations of this paper include: (1) aiming at the strong correlation between the coding structure and video content of HEVC, a motion target detection method based on motion intensity is proposed. Firstly, the motion vector and block partition information are extracted from the HEVC compressed bitstream, and the motion vector is de-noised and normalized, which makes the motion vector more stable and reliable. Then the motion intensity of each block is calculated with block partition information. Finally, by setting the adaptive threshold, the unit block with small motion intensity can be filtered out, and the moving object can be obtained. In this paper, the moving target detection based on motion intensity value is realized. (2) the current pixel direction detection algorithm for traffic accident detection is mainly based on the distance difference of centroid coordinates and time difference. In view of the high computational complexity and poor reliability of these algorithms, this paper proposes a multi-parameter fuzzy logic theory to detect the abnormal events of vehicles by analyzing the variation of the area and the intensity of motion in the vehicle area. Firstly, the breadth-first search algorithm is used to construct the connected region with the maximum coding unit block containing the motion intensity value, and the area of the connected region and its corresponding motion intensity value are extracted. Secondly, according to the area gradient information, different vehicle types and traffic conflict scenarios are distinguished. Finally, the input parameters of fuzzy logic theory are made up of the motion intensity value of each frame, the area gradient grade and the difference between the motion intensity and the average value. According to the membership function of the vehicle motion characteristics, the vehicle abnormal events can be judged in the current frame. Through the experimental verification of 25 sets of traffic accident videos, 22 groups of traffic accidents can be accurately detected, and the detection rate is 88%, which proves the validity and reliability of the proposed algorithm.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491.31;TP391.41

【参考文献】