动态背景下行人检测模块的设计与实现

发布时间：2018-04-30 00:25

本文选题：动态背景 + 摄像机运动　；参考：《延边大学》2017年硕士论文

【摘要】：随着信息技术的发展,视频文件的数量飞速增长。如何在海量的视频文件中检索出关键事件和关键目标具有重要的研究意义和广泛的应用价值。尤其是在冗余信息较多的交通监控和安防监控视频中进行事件和目标检测更具有重要的现实意义。但是,如何在复杂场景视频中实现快速准确的目标检测与行人检测,仍是目前运动目标检测领域亟待解决的问题。光流法是一种有效检测运动物体与运动行人的算法,不需任何先验知识,对背景模型不存在依赖,在运动目标检测领域有着广泛的应用。但光流法存在对噪声敏感、算法复杂度较高等缺陷,本学位论文针对光流法对噪声敏感和实时性差的缺点进行改进,提出了一种改进的光流法并将其应用于运动目标检测。本文主要研究内容如下:首先,针对摄像机运动造成的全局动态背景,采用改进的灰度投影法进行全局运动补偿。针对灰度投影法存在累积误差等缺陷,提出隔三帧更换一次参考帧的方法,减少了由于选择固定参考帧导致的累积计算误差;针对灰度投影法的投影区域存在运动目标时计算误差较大的问题,本学位论文使用视频图像帧的边角地带作为投影区域来计算运动矢量,大大减少了运动目标对运动补偿的影响。其次,提出了一种基于梯度阈值和特征抑制的光流法进行运动目标检测。将LK光流法和HS光流法进行结合,对光流约束方程进行改进,对梯度较大的像素点采用亮度约束,对梯度较小的像素点采用全局平滑约束,以确保光流约束方程的适用性。特征抑制作为辅助判断有效光流点的手段,在运动目标检测算法中对噪声和局部动态背景进行了有效地抑制。最后,在运动目标检测算法基础上,采用改进的形状复杂度特征,对视频中的运动目标进行分类。通过比较不同类型运动目标的形状复杂度取值范围,确定能够区分行人和其他类型运动目标的阈值,从而有效地检测出视频中的行人目标。测试结果表明,本文提出的动态背景下行人检测算法与其他算法相比,准确率和稳定性都有明显的提高。此外,本文提出的算法能够对摄像机运动、局部动态背景具有良好的鲁棒性,为下一步工作提供了良好的基础。
[Abstract]:With the development of information technology, the number of video files is increasing rapidly. How to retrieve the key events and key objects from a large number of video files has important research significance and wide application value. Especially, the detection of events and targets in traffic surveillance and security surveillance video with more redundant information is of great practical significance. However, how to achieve fast and accurate target detection and pedestrian detection in complex scene video is still an urgent problem in the field of moving target detection. Optical flow method is an effective algorithm for detecting moving objects and pedestrians. It does not require any prior knowledge and does not depend on the background model. It is widely used in the field of moving target detection. However, the optical flow method is sensitive to noise and the complexity of the algorithm is high. In this dissertation, an improved optical flow method is proposed and applied to the detection of moving targets. In view of the shortcomings of the optical flow method for noise sensitivity and poor real-time performance, an improved optical flow method is proposed and applied to the detection of moving targets. The main contents of this paper are as follows: firstly, an improved gray projection method is used to compensate the global motion of the camera. Aiming at the defects of gray projection method, such as accumulative error, the method of replacing reference frame once every three frames is put forward, which reduces the accumulated calculation error caused by selecting fixed reference frame. In view of the problem that the projection region of gray projection method has the problem of large calculation error when moving object, this thesis uses the edge zone of the video frame as the projection region to calculate the motion vector. The effect of moving target on motion compensation is greatly reduced. Secondly, an optical flow method based on gradient threshold and feature suppression is proposed to detect moving targets. The LK optical flow method and HS optical flow method are combined to improve the optical flow constraint equation. The brightness constraint is applied to the pixel points with large gradient and the global smoothing constraint is applied to the smaller gradient pixel points to ensure the applicability of the optical flow constraint equation. Feature suppression is used as an auxiliary method to judge the effective optical flow point. The noise and local dynamic background are effectively suppressed in the moving target detection algorithm. Finally, based on the moving target detection algorithm, the improved shape complexity feature is used to classify the moving target in video. By comparing the range of shape complexity of different types of moving objects, the threshold can be determined to distinguish pedestrians from other types of moving objects, so as to effectively detect pedestrian targets in video. The test results show that the proposed pedestrian detection algorithm in dynamic background is more accurate and more stable than other algorithms. In addition, the proposed algorithm is robust to camera motion and local dynamic background, and provides a good basis for further work.
【学位授予单位】：延边大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】